Google DeepMind
Key Information
- Formed: April 2023, through the merger of DeepMind Technologies (founded 2010) and Google Brain. [2, 35]
- Founders (DeepMind): Demis Hassabis, Shane Legg, Mustafa Suleyman. [2, 18]
- Headquarters: London, UK (with global research centres including USA, Canada, France, Germany, Switzerland). [2]
- Parent Company: Alphabet Inc. [2]
- Flagship Models: Gemini 3.8 Flash (released September 2, 2026; current strongest for reasoning, coding, agentic work). Gemini 3.5 Flash (May 2026). Gemini 3.5 Pro announced May 2026 but remains unreleased as of September 2026. Gemini 3.1 Pro, Gemma (open models), Veo (video). [2, 41]
- Main Products/Technologies: AlphaFold (protein folding), AlphaGo/AlphaZero (games), Imagen (text-to-image), Lyria (text-to-music), GNoME (materials science), Project Astra (universal AI assistant). [2, 28] Powers many Google products (Search, Cloud AI, Android, Vertex AI, Gemini App). [41]
- Official Website: deepmind.google [2]
- Research & Publications: Primarily via deepmind.google/research/publications/ and ai.google/research/pubs . [42, 43]
Origin & Structure
DeepMind Technologies was founded in London in 2010 with the goal to "solve intelligence." [2, 18, 35] Google acquired it in 2014. [2, 17, 26, 29] In April 2023, DeepMind merged with the Google Brain team to form Google DeepMind, a unified AI division within Alphabet Inc. [2, 28, 35]
Details
Key Milestones
- DeepMind Technologies (2010): Founded in London by Demis Hassabis, Shane Legg, and Mustafa Suleyman with the ambitious mission to understand and build artificial general intelligence. [2, 18, 28]
- Google Acquisition (2014): Acquired by Google for a reported sum between $400 million and $650 million, operating with considerable research autonomy. [2, 17, 26, 29, 33] An ethics board was part of the acquisition terms. [2]
- Google Brain: A separate, highly influential AI research team within Google, responsible for breakthroughs like TensorFlow and significant contributions to Transformer architectures. [2]
- Google DeepMind (April 2023): The formal consolidation of DeepMind and the Google Brain team, bringing together Google's AI research efforts under the leadership of Demis Hassabis as CEO of Google DeepMind, a subsidiary of Alphabet Inc. [2, 28, 35]
Philosophy & Approach
Google DeepMind pursues a science-led approach to AGI, emphasizing fundamental research and responsible AI development. [35] They aim to apply AI to solve major scientific and societal challenges, guided by Google's AI Principles. Explore their publications . [42]
Details
Core Beliefs & Strategy
- Solving Intelligence: A long-term, foundational commitment to understanding and building AGI. [35]
- Science & Research Driven: Strong emphasis on pioneering research, publishing extensively, and tackling grand scientific challenges like protein folding (AlphaFold), fusion energy control, and materials discovery (GNoME). [2, 28]
- Responsible Innovation: Adherence to Google's AI Principles, focusing on safety, ethics, fairness, transparency, and societal benefit. This includes a dedicated Responsibility & Safety team and ongoing ethics research. [2]
- Real-world Impact: Aims to translate AI breakthroughs into applications that benefit humanity, from scientific tools to enhancing Google's suite of products and services.
- Interdisciplinary Approach: Combines insights from machine learning, neuroscience, engineering, mathematics, and simulation. [28, 35]
Leadership
Led by co-founder and CEO Demis Hassabis. Lila Ibrahim serves as COO. [2] Koray Kavukcuoglu is CTO. [41]
Details
Key Figures (as of May 2025)
- Demis Hassabis: Co-founder and Chief Executive Officer (CEO) of Google DeepMind. Co-founder of Isomorphic Labs. Awarded the Nobel Prize in Chemistry 2024 for AlphaFold. [2]
- Lila Ibrahim: Chief Operating Officer (COO). [2]
- Koray Kavukcuoglu: Chief Technology Officer (CTO). [41]
- Co-founders Shane Legg remains with Google DeepMind. Mustafa Suleyman left in 2019, joined Google, and is now CEO of Microsoft AI as of March 2024. [2]
Key Models & Products/Technologies
Leading with Gemini 3.8 Flash (released September 2, 2026; strongest for reasoning, coding, agentic work). Earlier Gemini 3.5 Flash (May 2026). Gemini 3.5 Pro announced May 2026 but unreleased as of September 2026. Gemini 3.1 Pro, Gemma open models, Veo video generation. [2, 41] Renowned for AlphaFold (biology), AlphaGo/AlphaZero (games), Imagen (image generation), and Lyria (music generation). [2, 28] Explore more at Google DeepMind Technologies .
Details
Flagship Model Families
-
Gemini 3.8 (current flagship family, as of September 2026):
Google DeepMind's most capable and general multimodal model family, designed for text, code, image, audio, and video understanding and generation.
-
Gemini 3.8 Flash: Released September 2, 2026. Strongest agentic, coding, and reasoning model; optimized for inference speed; pricing at $0.75/$3.75 per million input/output tokens. [41] -
Gemini 3.5 Flash: Launched May 19, 2026 at Google I/O. 1M-token context; ~4Ă— faster than comparable frontier models at inference. -
Gemini 3.5 Pro: Announced May 19, 2026 at Google I/O for June 2026 availability, but remains unreleased as of September 2026 (multiple delays reported; no confirmed general availability date). [41] - Powers features in Google Search, Gemini App, Google Cloud AI (Vertex AI), Android, and Project Astra. [41]
-
- Gemma: A family of lightweight, state-of-the-art open models built from the same research and technology used for Gemini.
Groundbreaking AI Systems & Technologies
- AlphaFold: Revolutionized biology by accurately predicting 3D protein structures for nearly all known proteins, with data publicly available. [2, 28]
- AlphaGo / AlphaZero: AI systems that mastered complex board games like Go, chess, and shogi through self-play and reinforcement learning, defeating world champions. [2, 18]
- Imagen: Advanced text-to-image diffusion model series.
- Veo: High-quality text-to-video generation model; Veo 2 released Dec 2024. [2]
- Lyria: Text-to-music generation model, available in preview on Vertex AI. [2]
- GNoME (Graph Networks for Materials Exploration): AI tool that discovered millions of new stable crystalline materials. [2]
- Project Astra: Research initiative focused on building universal AI assistants with multimodal understanding and real-time interaction. [40, 41]
- Contributions to core AI technologies like Transformers and reinforcement learning.
Product Integration & Platforms
Google DeepMind's research and models are deeply integrated into Google's product ecosystem, including Google Search, Google Assistant, Google Photos, Google Workspace, Pixel devices, and provide foundational models for Google Cloud AI (Vertex AI). Follow their progress on the Google DeepMind Blog . [42]
AGI/ASI Goals & Approach
The foundational long-term research goal is to "solve intelligence," culminating in AGI. [2, 35] This is pursued through scientific breakthroughs, responsible development, and scaling general-purpose systems like Gemini.
Details
Approach to Advanced AI
- Long-term Aspiration: The original and ongoing mission is to achieve AGI. [2, 35] Demis Hassabis has suggested AGI could be developed within the next decade.
- Responsible & Safe AGI: A strong emphasis is placed on developing AGI safely and ethically, ensuring it is beneficial and controllable. This includes research into AI alignment, governance, and societal impact, guided by Google's AI Principles and a dedicated ethics team. [2]
- Pathways to AGI: Focus areas include reinforcement learning, neuroscience-inspired AI, large-scale multimodal modeling (e.g., Gemini), and developing more general and capable agentic systems (e.g., Project Astra, experimental agents in games with Gemini 2.0). [41]
- Scientific Application for Progress: Belief that tackling complex scientific problems (like AlphaFold for protein folding or GNoME for materials science) drives progress towards more general intelligence and demonstrates AI's potential benefits. [2, 28]
- Societal Readiness & Governance: Hassabis has expressed the need for societal preparedness for AGI and advocates for international cooperation and standards in AI development.
Funding & Resources
As a subsidiary of Alphabet Inc., Google DeepMind has access to Alphabet's extensive financial, computational (including Google's custom TPUs), and data resources. [2] The original DeepMind acquisition by Google in 2014 was reportedly $400-$650M. [2, 17, 26, 29]
Details
Resource Allocation
- Subsidiary of Alphabet: Benefits from Alphabet's significant R&D budget and infrastructure, including vast computing power (CPUs, GPUs, and Google's own Tensor Processing Units - TPUs) and large datasets. Specific internal budget allocations are not typically made public. [2]
- Original Acquisition Value: DeepMind Technologies was acquired by Google in 2014 for a sum reported to be between $400 million and $650 million. [2, 17, 26, 29, 33]
- Google.org Support: Google's philanthropic arm, Google.org, has committed funds (e.g., $20 million in Nov 2024) to support external academic and non-profit organizations using AI for science, often leveraging Google DeepMind's expertise.
- Isomorphic Labs: A sister company under Alphabet, also led by Demis Hassabis, focuses on AI for drug discovery, building on AlphaFold's success. It raised $600 million in external funding in early 2025.
Recent Developments (2024-2026)
Released Gemini 3.8 Flash (September 2, 2026) — strongest reasoning, coding, agentic model at $0.75/$3.75 per million tokens. Earlier Gemini 3.5 Flash (May 2026) with 1M-token context and 4× faster inference. Gemini 3.5 Pro (announced May 2026) remains unreleased as of September 2026 (multiple delays, no confirmed release date). [2, 41] Demis Hassabis awarded Nobel Prize in Chemistry (2024) for AlphaFold. [2] Continued Gemma open model releases. Veo 2 (Dec 2024) and Lyria music generation. [2, 41]
Details
Key Announcements & Progress
- Gemini Model Suite Evolution: Gemini 3.8 Flash released September 2, 2026 — strongest reasoning and coding model at $0.75/$3.75 per million input/output tokens. Earlier Gemini 3.5 Flash (May 2026) with 1M-token context and 4× faster inference than comparable frontier models. Gemini 3.5 Pro was announced May 2026 for June availability but remains unreleased as of September 2026 with multiple delays reported and no confirmed general availability date. Gemini 3.1 Pro remains available for complex reasoning. [41]
- Gemma Open Models: Continued development and release of Gemma, a family of lightweight, open models derived from Gemini research.
- Project Astra: Significant progress showcased on a universal AI assistant capable of real-time multimodal understanding and interaction. [40, 41]
- Nobel Prize Recognition: Demis Hassabis (CEO) and John Jumper (Senior Staff Research Scientist) were awarded the 2024 Nobel Prize in Chemistry for their groundbreaking work on AlphaFold. [2]
- AI for Science: Ongoing breakthroughs in applying AI to scientific discovery, including materials science (GNoME), weather forecasting, and fusion research. [2]
- Multimodal Generation: Release of Veo 2 (video generation, Dec 2024) and Lyria (text-to-music, available in preview on Vertex AI). [2]
- Responsible AI: Continued focus on AI safety, ethics, and governance, contributing to global discussions and standards.
- Isomorphic Labs Progress: Sister company Isomorphic Labs, leveraging DeepMind's AI for drug discovery, secured $600 million in external funding in early 2025.
Anthropic
Key Information
- Founded: 2021, by Dario Amodei, Daniela Amodei, Tom Brown, Chris Olah, Sam McCandlish, Jack Clark, Jared Kaplan, and others.
- Headquarters: San Francisco, California, USA.
- Valuation: ~$965 billion post-money after a $65 billion Series H (May 28, 2026), making Anthropic the most highly valued AI startup. Up from $380 billion (Series G, February 2026), $183 billion (September 2025), and $61.5 billion (May 2025).
- Flagship Models: Claude Mythos 5.1 / Fable 5.1 (released September 1, 2026; newest flagship). Earlier Claude Opus 5 (July 24, 2026), Claude Opus 4.8 (May 2026). Claude Sonnet 5 (June 30, 2026), Claude Sonnet 4.6 (February 2026). Claude Haiku 4.5. [Note: Mythos 5 initially released June 9, 2026 but suspended June 12-30 per US export controls; fully restored July 1, 2026.]
- Main Products: Claude.ai (chat interface and workspace), Anthropic API for developers, Claude models for enterprise.
- Official Website: anthropic.com
- Documentation: docs.anthropic.com
Origin & Founding Vision
Founded in 2021 by a group of former senior OpenAI researchers, including siblings Dario Amodei (CEO) and Daniela Amodei (President). Established as a Public Benefit Corporation (PBC) with a primary focus on AI safety and research.
Details
Key Details
- Founding Team: Composed of several ex-OpenAI leaders who shared concerns about the safety and societal impacts of increasingly powerful AI systems. Key founders include Dario Amodei, Daniela Amodei, Tom Brown, Chris Olah, Sam McCandlish, Jack Clark, and Jared Kaplan.
- Core Motivation: A desire to conduct AI research with an explicit and primary emphasis on safety, interpretability, and developing AI systems that are "helpful, honest, and harmless."
- Structure: Incorporated as a Public Benefit Corporation (PBC) to legally codify its commitment to public benefit and AI safety alongside its commercial objectives. Anthropic also has a unique "Long-Term Benefit Trust" designed to ensure its mission endures.
Philosophy: Safety-First AI
Anthropic is dedicated to building reliable, interpretable, and steerable AI systems. They have pioneered techniques like "Constitutional AI" and maintain a "Responsible Scaling Policy" to guide their development. See their research .
Details
Core Principles & Methodologies
- Helpful, Honest, and Harmless (HHH): These are the guiding desiderata for the behavior of their AI assistants.
- Constitutional AI: A methodology developed by Anthropic to train AI models based on a set of principles (a "constitution") derived from sources like the UN Universal Declaration of Human Rights. This aims to make AI behavior more aligned with human values and less reliant on extensive human labeling for harmful outputs.
- Responsible Scaling Policy (RSP): A framework outlining specific safety procedures and readiness levels (ASL-1, ASL-2, ASL-3 etc.) that must be met before developing or deploying more powerful AI models. This is intended to proactively manage risks as AI capabilities increase — see the AI Safety Ecosystem Hub for how this compares to other labs' frameworks.
- Interpretability Research: Significant research effort is dedicated to understanding the internal workings of large language models to make them more transparent, predictable, and trustworthy.
- Cautious and Iterative Deployment: Anthropic adopts a careful approach to deploying its models, aiming to learn from real-world interactions and continuously improve safety features.
Leadership
Co-founded and led by Dario Amodei (Chief Executive Officer) and Daniela Amodei (President). The leadership team includes many former senior members from OpenAI's safety and research divisions.
Details
Key Figures
- Dario Amodei: Co-founder and Chief Executive Officer (CEO). Formerly VP of Research at OpenAI.
- Daniela Amodei: Co-founder and President. Formerly VP of Safety and Policy at OpenAI.
- Other co-founders with significant roles include Tom Brown (key architect of GPT-3), Chris Olah (interpretability research lead), Jack Clark (policy and communications), Jared Kaplan (scaling laws research), and Sam McCandlish.
Key Models & Products
The Claude family of large language models is Anthropic's flagship offering. The current lineup (as of July 26, 2026): Claude Opus 5 (July 24, 2026; near-frontier performance at half Fable input price), Claude Opus 4.8 (May 2026; complex reasoning, long-horizon agentic coding), Claude Sonnet 5 (June 30, 2026; most agentic Sonnet), Claude Sonnet 4.6 (February 2026), and Claude Haiku 4.5 — see the model picker for which tier fits which task, and AI coding agents compared for how Claude Code stacks up against Cursor, Copilot, and Codex. Claude Fable 5 / Mythos 5 (released June 9, 2026, suspended June 12-30 per US government export control directive; both fully restored July 1, 2026) is now available again. These models are known for strong performance, long context windows, and safety features. Products include the Claude.ai chat interface and the Anthropic API for developers and enterprises.
Details
Claude Model Family
- Claude Fable 5 & Mythos 5 (Released June 9, 2026; Suspended June 12-30, 2026; Fully Restored July 1, 2026): Anthropic's most capable widely released models at launch. First Mythos-class models made generally available, featuring always-on adaptive thinking, a 1M-token context window, and 128K output tokens, with state-of-the-art results on nearly all tested benchmarks. Were priced at $10 input / $50 output per million tokens. On June 12, 2026 at 5:21pm ET, the US government issued an export control directive citing national security concerns regarding a potential jailbreak method, requiring Anthropic to suspend all access to Fable 5 and Mythos 5 for all users (including foreign nationals and Anthropic's own foreign national employees). All users lost access on that date. On June 30, 2026, the US government lifted the export controls, and Anthropic fully restored both Claude Fable 5 and Claude Mythos 5 on July 1, 2026, globally on the Claude Platform, Claude.ai, Claude Code, and Claude Cowork. Anthropic had trained an improved safety classifier targeting the reported jailbreak technique, which now blocks the behavior in over 99% of attempts. Other Claude models (Opus 4.8, Sonnet 4.6, Haiku 4.5) remained available without restriction throughout.
- Claude Opus 4.8 (Released May 2026): Strongest Opus-tier model; excels at complex reasoning, long-horizon agentic coding, and high-autonomy tasks. Priced at $5 input / $25 output per million tokens.
- Claude Sonnet 4.6 (Released February 2026): Balanced speed and capability; the model Claude.ai uses for Pro subscribers. Priced at $3 input / $15 output per million tokens.
- Claude Haiku 4.5: Fastest and most cost-effective model ($1 input / $5 output per million tokens), suited for high-throughput, latency-sensitive applications.
Key Products & Platforms
- Claude.ai : Web-based chat interface and workspace for interacting with Claude models, offering free and paid tiers (Claude Pro). Includes features like Artifacts for dynamic content.
- Anthropic API : Provides developer access to the Claude model family for integration into custom applications and services. Documentation available at docs.anthropic.com .
- Enterprise Offerings: Tailored solutions and model access for businesses, emphasizing safety, reliability, and customization.
- Cloud Partnerships: Claude models are available on major cloud platforms, including Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Azure, expanding accessibility for enterprises.
AGI/ASI Goals & Safety
Anthropic views AGI development as a serious undertaking requiring proactive and deeply integrated safety measures. Their goal is to ensure that advanced AI systems are beneficial and steerable, with safety research informing every stage of development.
Details
Approach to Advanced AI
- Safety-Centric AGI Development: While aiming to build highly capable AI, Anthropic's primary differentiator is the profound integration of safety research and principles (like Constitutional AI) directly into the model development process from the outset.
- Proactive Risk Mitigation (RSP): Their Responsible Scaling Policy (RSP) is a public commitment to a staged approach for developing increasingly powerful models, with specific safety measures and evaluations required at each AI Safety Level (ASL).
- Steerable and Interpretable AI: A core research focus is on making AI models more understandable (interpretability) and controllable (steerability), so their behavior can be reliably guided by human intentions and ethical principles.
- Long-Term Benefit & Governance: The overarching goal is to ensure that future AGI systems serve humanity's long-term interests and avoid harmful outcomes. This includes considerations for governance structures, such as their Long-Term Benefit Trust.
Funding & Investors
Anthropic has secured tens of billions in funding from Google, Amazon, and major venture firms. A $30 billion Series G (February 2026) set a $380 billion valuation. A $65 billion Series H (May 28, 2026) set a $965 billion post-money valuation — surpassing OpenAI and potentially Anthropic's final private raise before IPO. Run-rate revenue crossed $47 billion by Series H.
Details
Key Investments & Valuation
- Google: A significant investor, with initial investments and commitments reportedly up to $2 billion, and an additional $550 million reported. Google Cloud is a key partner.
- Amazon: Committed up to $4 billion, making AWS Anthropic's primary cloud provider for mission-critical workloads. Amazon Bedrock offers Claude models.
- Microsoft: Reported commitment of $2 billion, with Claude models also available on Azure.
- Other Key Investors: Include Spark Capital, Salesforce Ventures, Sound Ventures, Menlo Ventures, SK Telecom, Lightspeed Venture Partners, General Catalyst, Jane Street, and Fidelity.
- Total Funding Secured: Cumulative equity and committed funding has grown into the hundreds of billions, including a $30 billion Series G (February 2026) and a $65 billion Series H (May 28, 2026). Run-rate revenue was $14 billion at Series G and crossed $47 billion by Series H.
- Valuation Trajectory: Climbed from $15–18.4 billion (late 2023/early 2024) to $61.5 billion (May 2025), $183 billion (Series F, September 2025), $380 billion (Series G, February 2026), and $965 billion (Series H, May 28, 2026). At $965 billion, Anthropic briefly surpassed OpenAI as the most highly valued AI startup. Series H led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital; likely the final private round before IPO.
Recent Developments (2024-2026)
Released Claude Mythos 5.1 / Fable 5.1 (September 1, 2026 — newest flagship, most advanced for coding and knowledge work). Earlier launches: Claude Opus 5 (July 24, 2026 — near-frontier at half Fable input price), Claude Sonnet 5 (June 30, 2026 — near-Opus performance), and Fable 5 / Mythos 5 (June 9, 2026; suspended June 12-30 by US export controls; restored July 1, 2026). Closed $30B Series G at $380B (February 2026) and $65B Series H at $965B (May 28, 2026). Run-rate revenue crossed $47B. Check their news page .
Details
Key Announcements
- Claude Sonnet 5 (Released June 30, 2026): Latest Sonnet-tier model; positioned as "most agentic Sonnet" with near-Opus performance at Sonnet pricing. Became default on Claude.ai same day. Introductory pricing $2/$10 per million input/output tokens through August 31, 2026; standard pricing $3/$15 afterward.
- Claude Opus 5 (Released July 24, 2026): Latest Opus-tier model; "near-frontier default" approaching Claude Fable 5 performance at roughly half input price. Strong on SWE-bench Pro (79.2%), everyday coding, agentic work, knowledge tasks. Pricing roughly 1.7x Claude Sonnet 5.
- Claude Opus 4.8 (May 2026): Prior strongest Opus; complex reasoning and long-horizon agentic coding with high autonomy. $5/$25 per million input/output tokens.
- Claude Sonnet 4.6 (February 2026): Prior Sonnet flagship; full upgrade across coding, computer use, long-context reasoning, agent planning, knowledge work, and design. Prior default for Claude.ai Pro subscribers.
- Claude Opus 4.8 (May 2026): Strongest Opus-tier model; complex reasoning and long-horizon agentic coding with high autonomy. $5/$25 per million input/output tokens.
- Claude Fable 5 & Mythos 5 (June 9, 2026 — Suspended June 12-30; Fully Restored July 1): First Mythos-class models released publicly, featuring always-on adaptive thinking, 1M-token context window, and 128K output tokens; achieved state-of-the-art on nearly all benchmarks. Were priced at $10/$50 per million input/output tokens and available on Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Azure. On June 12, 2026 at 5:21pm (ET), the US government issued an export control directive citing national security concerns regarding a potential jailbreak method discovered by Amazon researchers, requiring Anthropic to suspend all access to both Fable 5 and Mythos 5 for all users, including foreign nationals and Anthropic employees. All access was terminated on that date. On June 30, 2026, the US government lifted the export controls. Anthropic fully restored both Claude Fable 5 and Claude Mythos 5 on July 1, 2026, globally on Claude Platform, Claude.ai, Claude Code, and Claude Cowork. Anthropic trained an improved safety classifier targeting the reported jailbreak technique, now blocking the behavior in over 99% of attempts. Other Claude models (Opus 4.8, Sonnet 4.6, Haiku 4.5) remained unaffected throughout.
- Series H Funding (May 28, 2026): Raised $65 billion at a $965 billion post-money valuation, led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital. Surpassed OpenAI as most highly valued AI startup. Run-rate revenue crossed $47 billion. Likely Anthropic's final private fundraise before IPO.
- Series G Funding (February 2026): Raised $30 billion at a $380 billion post-money valuation, with run-rate revenue of $14 billion at that time.
- Enterprise Expansion & Cloud Availability: Deep partnerships with AWS (primary cloud), Google Cloud Vertex AI, and Microsoft Azure Foundry for enterprise model access.
- Responsible Scaling Policy (RSP) Updates: Continued commitment and updates to their RSP, detailing safety levels and procedures for developing more advanced AI.
- Research Publications: Ongoing release of influential research papers on AI safety, interpretability, and model capabilities at anthropic.com/research .
OpenAI
Key Information
- Founded: December 2015, by Elon Musk, Sam Altman, Greg Brockman, Ilya Sutskever, Wojciech Zaremba, John Schulman, and others. [1]
- Headquarters: San Francisco, California, USA. [1]
- Valuation: ~$852 billion post-money valuation after a record $122 billion funding round (March 2026). Up from $300 billion (April 2025) and $157 billion (October 2024). IPO speculation points toward a possible 2027 listing near $1 trillion.
- Flagship Models: GPT-6 Astra (released September 3, 2026; $10/$50 per 1M tokens; current flagship), GPT-6 Sol ($2/$10) and GPT-6 Luna ($0.10/$0.50), both released September 22, 2026 at half the price of their GPT-5.6 predecessors. GPT-5.6 series (Sol, Terra, Luna — July 9, 2026 preview, ~1.05M-token context; superseded). GPT-5.5 (April 2026), GPT-5.4 (March 2026). DALL-E 3, Sora, Whisper, reasoning capabilities. Deep Research. Earlier GPT-4 family (retired Feb 13, 2026) and o1 have been retired. [1, 11]
- Main Products: ChatGPT (various tiers), OpenAI API, specialized models for enterprise.
- Official Website: openai.com [1]
- Documentation: platform.openai.com/docs [11]
Origin & Founding Vision
Founded in December 2015 as a non-profit research organization, OpenAI later adopted a "capped-profit" model to attract investment for large-scale AI research. [1] Its core mission is to ensure that artificial general intelligence (AGI) benefits all of humanity. Learn more on their about page .
Details
Key Details
- Founding Goal: To build Artificial General Intelligence (AGI) that is safe and broadly beneficial, as outlined in their charter. [1]
- Initial Structure: Non-profit research company (OpenAI, Inc.). [1]
- Key Founders: Included notable figures such as Elon Musk, Sam Altman, Greg Brockman, Ilya Sutskever, Wojciech Zaremba, and John Schulman. [1]
- Transition to "Capped-Profit": In 2019, OpenAI LP was formed as a capped-profit subsidiary to raise the substantial capital needed for compute-intensive research, while the non-profit OpenAI, Inc. remains the overall governing body with its mission as primary. [1, 12]
- Current Structure (as of 2025): A complex structure involving the non-profit OpenAI, Inc. and for-profit subsidiaries like OpenAI Global, LLC, which handles commercial operations. [1] Microsoft has a significant partnership, providing funding and Azure cloud resources, and is entitled to a share of OpenAI Global, LLC's profits. [1, 12, 14]
Philosophy & Culture
OpenAI's philosophy centers on ambitious research towards AGI, coupled with a strong emphasis on safety, responsibility, and ensuring broad societal benefit. [1] They advocate for iterative deployment of increasingly powerful AI systems to foster societal adaptation and learning. Read their research . [11]
Details
Core Tenets
- Beneficial AGI: The primary mission is to ensure that AGI, defined as highly autonomous systems outperforming humans at most economically valuable work, benefits all of humanity. [1]
- Safety Research & Preparedness: Significant investment in AI safety research to mitigate risks from powerful AI. [13] They developed a "Preparedness Framework" to assess and manage catastrophic risks associated with frontier AI models.
- Long-term Perspective: Acknowledges that AGI development is a long and challenging endeavor requiring sustained research efforts.
- Iterative Deployment: Believes in deploying increasingly capable AI systems to learn from real-world applications, allowing society to adapt and for safety measures to be refined based on empirical evidence.
- Evolving Openness: While initially having a strong open-source ethos, OpenAI has become more selective about releasing its most powerful models, citing safety and competitive reasons. However, it continues to publish research and release some models and tools (e.g., on GitHub ).
Leadership
Led by CEO Sam Altman and President Greg Brockman. [16] Mira Murati departed as CTO in September 2024; Mark Chen leads research as SVP of Research. The board of the non-profit OpenAI, Inc. is chaired by Bret Taylor. Fidji Simo serves as CEO of Applications (from 2025). [1, 15]
Details
Key Figures (as of June 2026)
- Sam Altman: Chief Executive Officer (CEO) of OpenAI. [1, 16]
- Greg Brockman: President and Co-founder. [1, 16]
- Mark Chen: SVP of Research, leading the research organization after Mira Murati's departure (September 2024). [13]
- Brad Lightcap: Chief Operating Officer (COO). [13, 16]
- Sarah Friar: Chief Financial Officer (CFO). [1]
- Fidji Simo: CEO of Applications. [15]
- Julia Villagra: Chief People Officer. [13]
- Bret Taylor: Chairman of the Board of Directors (OpenAI, Inc. nonprofit). [1]
- Former NSA Director Paul Nakasone joined the board in June 2024. Mira Murati (former CTO) departed September 2024 to found Thinking Machines Lab.
Note: OpenAI underwent a significant leadership event in November 2023, with Altman's brief removal and subsequent reinstatement. [1] The leadership structure continues to evolve as the company scales. [15, 32]
Key Models & Products
Known for GPT-6 Astra (released September 3, 2026; $10/$50 per 1M tokens; current flagship), GPT-6 Sol ($2/$10) and GPT-6 Luna ($0.10/$0.50), both released September 22, 2026 at half the price of their GPT-5.6 predecessors. Earlier GPT-5.6 series (Sol, Terra, Luna — July 9, 2026 preview; ~1.05M-token context), GPT-5.5 (April 2026), GPT-5.4 (March 2026). DALL-E 3 (image generation), Sora (text-to-video), Whisper (speech-to-text), and Deep Research agent. Earlier GPT-4 family (retired Feb 13, 2026) and o1 have been retired. [1] Products include ChatGPT (free, Plus, Team, Enterprise), and the OpenAI API for developers. [11]
Details
Prominent AI Models
-
GPT-6 (Astra released September 3, 2026; Sol and Luna released September 22, 2026):
-
GPT-6 Astra: Released September 3, 2026. Flagship model at $10/$50 per 1M input/output tokens. Reasoning and agentic capabilities. -
GPT-6 Sol: Released September 22, 2026 at $2/$10 per 1M input/output tokens; built for complex coding and agentic workflows; 1.05M-token context. -
GPT-6 Luna: Released September 22, 2026 at $0.10/$0.50 per 1M input/output tokens; most efficient model for focused, high-volume tasks; 1.05M-token context. - GPT-5.6 Series (July 2026 preview): Prior line with Sol, Terra, Luna variants. ~1.05M-token context, released July 9, 2026. GPT-6 Sol and Luna replaced Sol and Luna at half the price; there is no GPT-6 Terra.
-
GPT-5.6 Terra: Balanced intelligence and cost; ~1.05M-token context, released July 9, 2026. Prior generation; GPT-6 Sol ($2/$10) now occupies the workhorse tier. -
GPT-5.6 Luna: Prior-generation cost tier; ~1.05M-token context, released July 9, 2026. GPT-6 Luna ($0.10/$0.50) replaced it at half the price. -
GPT-5.5: Released April 23, 2026 as prior flagship. Still available; described as OpenAI's "smartest and most intuitive" model for complex agentic tasks. -
GPT-5.4: Released March 2026; 1M-token context, native computer-use. Still available. - Note: The entire pre-GPT-5 model family (GPT-4, GPT-4o retired Feb 13 2026, GPT-4.1) has been retired.
-
-
o-Series (Reasoning Models):
-
o3&o4-mini: Advanced reasoning models; o4-mini has been retired from ChatGPT but reasoning capabilities are integrated into the GPT-5.x line. [1]
-
- DALL-E 3: Advanced AI system creating realistic images and art from natural language descriptions. [1]
- Sora: Text-to-video model capable of generating realistic and imaginative video scenes. [1, 11] Access expanded to ChatGPT Plus/Pro users (late 2024).
- Whisper: Versatile speech recognition (ASR) and translation model. [1]
- Deep Research: An agent leveraging o3 for extensive web browsing, data analysis, and report synthesis. [1]
Key Products & Platforms
- ChatGPT : Conversational AI interface available in free, Plus, Team, and Enterprise tiers, offering access to various models. [11]
- OpenAI API : Allows developers to integrate OpenAI's models into their own applications and services. Includes tools like the Responses API and Agents SDK for building AI agents (announced March 2025). [11]
- Specialized Enterprise Solutions: Tailored offerings for business customers.
- Partnerships: Strategic collaborations, notably with Microsoft for Azure cloud services and distribution [1, 12, 20], and Apple for integrating ChatGPT into Apple Intelligence (announced June 2024).
AGI/ASI Goals & Approach
OpenAI explicitly aims to build Artificial General Intelligence (AGI) that is safe and benefits all of humanity. [1] Their approach involves scaling deep learning models, iterative deployment, and dedicated safety research.
Details
Stated Ambition & Strategy
- Core Mission: The development of AGI is central to OpenAI's charter. [1] They define AGI as "highly autonomous systems that outperform humans at most economically valuable work." [1]
- Safety as a Priority: AGI development is pursued with a strong emphasis on alignment with human values and intentions. [13] OpenAI has a "Preparedness Framework" to evaluate and mitigate catastrophic risks from advanced AI.
- Path to AGI: Primarily involves scaling current deep learning architectures (like Transformers), complemented by research into new architectures, algorithms, and continuous safety improvements. Iterative deployment of increasingly capable systems is a key part of this strategy. [13]
- ASI Considerations: OpenAI acknowledges the potential for Artificial Superintelligence (ASI) beyond AGI and the profound societal implications, stressing the need for careful governance and global cooperation.
Funding & Valuation
Major financial backing from Microsoft, reportedly totaling around $13 billion. [1, 12, 20] In March 2026, OpenAI closed a record $122 billion funding round at an ~$852 billion post-money valuation, with major participation from SoftBank, NVIDIA, and Amazon. This followed a $300 billion valuation (April 2025) and $157 billion (October 2024).
Details
Key Investments & Financials
- Microsoft Partnership: A multi-year, multi-billion dollar investment (around $13 billion reported) providing crucial funding and Azure cloud computing resources. Microsoft is entitled to a significant share of profits from OpenAI's for-profit arm. [1, 12, 14, 20]
- March 2026 Funding Round: Closed a record $122 billion round at an ~$852 billion post-money valuation, the largest venture deal ever. Major backers included SoftBank, NVIDIA, and Amazon. This followed an April 2025 round of $40 billion led by SoftBank at a $300 billion valuation.
- October 2024 Valuation: Valued at $157 billion during a previous funding phase. [8]
- Projected Revenue & Costs: Run-rate revenue reached roughly $24 billion annualized by early 2026 (~$2 billion/month), up from an estimated $3.7 billion in 2024. [1] However, compute costs remain substantial, with tens of billions in projected annual spending. [6]
- Early Backers: Initial support came from Sam Altman, Greg Brockman, Elon Musk, Reid Hoffman, Peter Thiel, and others. [1]
- Stargate Project: A significant portion of new funding is reportedly allocated to "Stargate," a joint supercomputer project with SoftBank and Oracle. [10]
Recent Developments (2024-2026)
Released GPT-6 Astra (September 3, 2026, new flagship) and GPT-6 Sol and Luna (September 22, 2026, at half the GPT-5.6 price). Earlier GPT-5.6 series (Sol, Terra, Luna — July 9, 2026 preview, ~1.05M-token context), GPT-5.5 (April 2026), GPT-5.4 (March 2026). Entire pre-GPT-5 model family (GPT-4 retired Feb 13, 2026, and earlier) retired. [1, 11] Closed record $122B funding round at an ~$852B valuation (March 2026). Stay updated via their blog . [11]
Details
Key Announcements & Activities
- Model Releases & Enhancements: GPT-6 Astra launched September 3, 2026, and GPT-6 Sol and Luna on September 22, 2026 (all with ~1.05M-token context; Sol $2/$10, Luna $0.10/$0.50). GPT-5.6 series launched July 9, 2026 (Sol, Terra, Luna with ~1.05M-token context, Feb 16 knowledge cutoff) and is now superseded. GPT-5.5 launched April 23, 2026 as prior flagship. GPT-5.4 launched March 2026 with 1M-token context and native computer-use. GPT-5.5 Instant available free from May 2026 as default for free users. The entire pre-GPT-5 model family (GPT-4, GPT-4o retired Feb 13, 2026, GPT-4.1, o1) has been retired. Sora text-to-video model access expanded. Deep Research agent (Feb 2025) and Responses API / Agents SDK (March 2025) remain available. [1, 11]
- Developer Tools: Responses API and Agents SDK (announced March 2025) for building AI agents; GPT-5.5 and GPT-5.5 Pro available in the API from April 24, 2026.
- Partnerships & Integrations: Integration of ChatGPT into Apple Intelligence (announced June 2024). Ongoing strong partnership with Microsoft Azure. [1, 12, 20] Agreement with CoreWeave for AI infrastructure (March 2025). [1]
- Funding & Corporate: Closed a record $122 billion funding round at an ~$852 billion post-money valuation (March 31, 2026), anchored by Amazon ($50B), Nvidia ($30B), and SoftBank ($30B), following the $40 billion / $300 billion round (April 2025). IPO speculation points toward a possible 2027 listing. [14]
- Leadership & Board: CTO Mira Murati departed September 2024; Mark Chen became SVP of Research. Fidji Simo serves as CEO of Applications (from 2025). [15] Former NSA Director Paul Nakasone joined the Board of Directors (June 2024). [13]
- Safety Framework: Continued updates to its Preparedness Framework for assessing and mitigating AI risks.
xAI
Key Information
- Founded: 2023, by Elon Musk with a founding team of researchers from DeepMind, OpenAI, Google, Microsoft, and Tesla.
- Headquarters: Bay Area, California, USA, with its "Colossus" supercomputer cluster in Memphis, Tennessee.
- Valuation: ~$230–250 billion. A $20 billion Series E (January 2026), led by Nvidia, was followed by SpaceX absorbing xAI in a deal valuing the combined entity around $1.25 trillion.
- Flagship Models: Grok 4.6 (released August 12, 2026 — current flagship; Grok 4.7 announced but not yet released as of September 12, 2026). Earlier Grok 4.5 (July 8, 2026; 500K context window, native video input). Grok 4.3 (Beta, April 2026). Grok 3 (released Feb 17, 2025; retired August 15, 2026). Grok-1 weights released openly. Strong focus on reasoning and real-time knowledge from X.
- Main Products: Grok conversational assistant (integrated into the X platform and Tesla vehicles), Grok API, and standalone Grok apps.
- Official Website: x.ai
- Documentation: docs.x.ai
Origin & Focus
Founded by Elon Musk in 2023 as a counterweight to what he viewed as overly cautious or ideologically biased AI labs. xAI's stated purpose is to build a "maximally truth-seeking" AI to help humanity understand the universe, leveraging tight integration with X (formerly Twitter) for real-time data.
Details
Key Details
- Founding Context: Musk, an early backer of OpenAI, launched xAI after publicly criticizing the direction of leading labs. The team was assembled from veterans of DeepMind, OpenAI, Google Research, and Tesla's Autopilot/AI teams.
- Mission: "Understand the true nature of the universe." xAI frames its goal as building curious, truth-seeking AI rather than systems optimized purely for engagement or safety-by-restriction.
- Ecosystem Integration: Deep ties to Musk's other companies — real-time data from X, distribution through the X app and Tesla vehicles, and (following the 2026 merger) compute and capital alignment with SpaceX.
Philosophy & Approach
xAI emphasizes "truth-seeking" over heavy content moderation, rapid iteration, and brute-force scaling of compute. It positions Grok as a less filtered, more candid assistant, and pursues an aggressive build-out of in-house supercomputing capacity.
Details
Core Principles
- Truth-Seeking AI: xAI argues that AI should aim to be maximally accurate and curious, and is openly skeptical of what it characterizes as excessive filtering in rival models.
- Compute-First Scaling: The "Colossus" cluster in Memphis was stood up unusually quickly and scaled to hundreds of thousands of GPUs, reflecting a belief that raw compute is a primary driver of capability.
- Speed of Iteration: xAI has shipped successive Grok generations on a compressed timeline, prioritizing fast public releases and frequent updates.
- Selective Openness: Released Grok-1 weights openly, though later flagship models are offered primarily as products and via API.
Leadership
Founded and led by Elon Musk, who also runs Tesla and SpaceX. The research organization is built from senior scientists and engineers recruited from DeepMind, OpenAI, and Google.
Details
Key Figures
- Elon Musk: Founder and CEO. Sets the company's direction and drives its capital, compute, and distribution strategy via SpaceX, Tesla, and X.
- Founding Research Team: A small, senior group drawn from DeepMind, OpenAI, Google Research, and Tesla AI, responsible for the Grok model line.
- SpaceX Integration: Following the 2026 merger, xAI's AI efforts (Grok and X) operate under a combined SpaceX/xAI structure, aligning leadership across compute and capital.
Key Models & Products
The Grok family has progressed rapidly from Grok 1 to Grok 4.5 (July 8, 2026; 500K token context, native video), adding reasoning, multimodal capability, and real-time knowledge from X. Grok is delivered through the X platform, standalone apps, Tesla vehicles, and the xAI API .
Details
Model Line
- Grok 4.6 (Released August 12, 2026): Current flagship. Grok 4.7 (~2.1T parameters, per xAI) has been announced but not released as of September 12, 2026.
-
Grok 4.5(Released July 8, 2026): Earlier flagship; 500K token context window, native video input. Trained on Colossus cluster. - Grok 4.3 (Beta, April 17, 2026): Improved architecture, native video input, document generation (PDFs, spreadsheets, slides), better tool-calling than earlier versions. December 2025 knowledge cutoff.
- Grok 3 (Deprecated): Released February 17, 2025 on Colossus cluster with 10x compute of Grok-2. Deprecated May 15, 2026; retiring August 15, 2026.
- Grok 1 / 1.5 / Grok 2: Earlier generations. Grok-1's weights published openly. Grok 2 brought competitive general performance and image generation in X platform.
Products & Platforms
- Grok Assistant: Conversational AI embedded in the X app, available as standalone apps, and rolled out into Tesla vehicles.
- xAI API: Developer access to Grok models for third-party applications.
- Colossus Supercomputer: A massive GPU cluster in Memphis, Tennessee, scaled to hundreds of thousands of accelerators to train successive Grok generations.
AGI/ASI Goals & Approach
xAI explicitly pursues AGI, framing the goal as a curious, truth-seeking intelligence that helps humanity understand the universe. Its bet is that massive compute scaling, fast iteration, and real-world data will be the primary path to general intelligence.
Details
Stated Ambition & Strategy
- Core Mission: To build AGI that is "maximally truth-seeking" and curious, which Musk argues is the safest long-term design because such a system would value understanding reality accurately.
- Path to AGI: Aggressive scaling of compute (the Colossus cluster), rapid model iteration, and grounding in real-time data from X.
- Safety Stance: xAI argues that curiosity and truthfulness are better safety properties than restriction-heavy alignment, a position that draws debate within the wider AI-safety community.
Funding & Investors
xAI has raised tens of billions in a short period, including a $20 billion Series E in January 2026 led by Nvidia with participation from Cisco, Fidelity, and others. In early 2026, SpaceX absorbed xAI in a deal valuing the combined entity around $1.25 trillion (xAI itself at roughly $250 billion).
Details
Key Investment Activity
- Early Rounds (2024): xAI raised multiple multi-billion-dollar rounds from major venture and strategic investors to fund its compute build-out.
- Series E (January 2026): Raised $20 billion — exceeding its initial target — in a round led by Nvidia, with Cisco, Fidelity, and other investors participating, at a post-money valuation in the $230 billion+ range.
- SpaceX Merger (2026): SpaceX absorbed xAI in a deal valuing the combined company around $1.25 trillion, aligning capital, compute, and leadership across Musk's space and AI businesses.
Recent Developments (2024-2026)
Shipped successive Grok generations through Grok 4, scaled the Colossus supercomputer in Memphis, raised a $20 billion Series E led by Nvidia, and was absorbed by SpaceX into a combined ~$1.25 trillion entity. Grok was rolled out across X and Tesla vehicles.
Details
Key Announcements & Activities
- Rapid Model Cadence: Released Grok 2, Grok 3, and Grok 4 in quick succession, steadily closing the gap with frontier models from OpenAI, Anthropic, and Google.
- Colossus Scale-Up: Continued expanding the Memphis supercomputer to hundreds of thousands of GPUs, one of the largest known training clusters.
- $20B Series E (January 2026): Closed a $20 billion round led by Nvidia, cementing xAI as one of the best-capitalized AI labs.
- SpaceX Merger (2026): SpaceX absorbed xAI; Musk announced Grok and X would operate under a combined SpaceX/xAI AI division.
- Distribution Expansion: Grok was integrated more deeply into the X platform and rolled out to Tesla vehicles, broadening its consumer reach.
DeepSeek
Key Information
- Founded: 2023, by Liang Wenfeng, spun out of the Chinese quantitative hedge fund High-Flyer.
- Headquarters: Hangzhou, China.
- Valuation: Self-funded by High-Flyer until 2026. Closed first external funding round in 2026, raising ~$7.4 billion (50B+ yuan) led by Tencent (~$1.5B) and CATL (~$735M), with Liang Wenfeng personally investing ~$2.8B. Post-money valuation reported at $52–59 billion.
- Flagship Models: DeepSeek-V4.1-Flash (released September 10, 2026 — current flagship). DeepSeek-V4 (April 2026). Earlier DeepSeek-V3, V2 (Mixture-of-Experts models) and DeepSeek-R1 (reasoning model), released with open weights and technical reports.
- Main Products: DeepSeek chatbot apps, a low-cost developer API, and openly downloadable model weights.
- Official Website: deepseek.com
- Documentation: api-docs.deepseek.com
Origin & Focus
DeepSeek grew out of High-Flyer, a hedge fund that had accumulated large GPU clusters for quantitative trading. Founder Liang Wenfeng redirected that compute and talent toward fundamental AI research, with a focus on training highly capable models efficiently and releasing them openly.
Details
Key Details
- High-Flyer Roots: The parent hedge fund had already built substantial GPU infrastructure, giving DeepSeek an unusual amount of compute for a young, self-funded lab.
- Mission: To pursue AGI through fundamental research, prioritizing open publication of models and methods over near-term commercialization.
- Breakout Moment: The January 2025 release of DeepSeek-R1 — a strong reasoning model trained at a fraction of typical cost — triggered a sharp global market reaction and intense scrutiny of frontier-lab spending.
Philosophy & Approach
DeepSeek emphasizes research efficiency, architectural innovation, and openness. It releases competitive models with permissive licenses and detailed papers, arguing that capability per dollar — not just raw spend — is the key frontier metric.
Details
Core Principles
- Efficiency-First Research: Heavy investment in Mixture-of-Experts architectures, training optimizations, and inference cost reduction to achieve frontier-level results with comparatively modest budgets.
- Open Weights & Open Research: Models such as V3 and R1 are released under permissive licenses with thorough technical reports, making DeepSeek one of the most influential open-model labs.
- Low-Cost Access: Its API is priced aggressively, pressuring the broader market on inference economics.
- Talent & Curiosity: A research culture built around a relatively small team of strong researchers, with funding raised in part to retain talent against aggressive poaching.
Leadership
Led by founder and CEO Liang Wenfeng, who also founded and runs the High-Flyer hedge fund. The research team is small, young, and recruited largely from top Chinese universities.
Details
Key Figures
- Liang Wenfeng: Founder and CEO of both DeepSeek and the High-Flyer hedge fund. Sets the lab's research-first, open-publication strategy.
- Research Team: A compact group of researchers and engineers, many recruited directly from leading Chinese universities, known for rapid iteration and strong publication output.
- High-Flyer Backing: The hedge fund provided early capital and compute, allowing DeepSeek to operate without external investors until 2026.
Key Models & Products
DeepSeek's model line spans efficient general-purpose models (V4 Pro/Flash — April 24, 2026; V3.2, V3 and V2) and the R1 reasoning series, all with open weights. They are available through DeepSeek's apps and a low-cost API , as well as direct download.
Details
Model Line
- DeepSeek-V4.1-Flash (Released September 10, 2026 — Current Flagship): Latest frontier model. Images and text input; new architecture and billing.
-
DeepSeek-V4(Released April 24, 2026): General-purpose model family. DeepSeek-V4-Pro and DeepSeek-V4-Flash variants; replaced V3.2 on main API routes. - DeepSeek-V3.2 (Released Dec 1, 2025): Previous flagship with integrated thinking in tool-use; open weights available.
- DeepSeek-V3, V2: Efficient Mixture-of-Experts models delivering competitive general performance with notable training and inference cost efficiency.
- DeepSeek-R1: An open reasoning model released in January 2025 that matched leading proprietary reasoning systems on many benchmarks, drawing intense global attention.
- Specialized Models: Additional open releases targeting coding and math, distributed with permissive licenses and technical reports.
Products & Platforms
- DeepSeek App: Consumer chatbot apps that briefly topped app-store charts following the R1 release.
- DeepSeek API: Developer access to its models at aggressively low pricing.
- Open Weights: Downloadable model weights that have made DeepSeek a foundation for a large ecosystem of derivative models.
AGI/ASI Goals & Approach
DeepSeek states that AGI is its long-term goal, pursued through fundamental research rather than rapid productization. Its distinctive bet is that algorithmic and architectural efficiency — not just scale of spend — is central to reaching general intelligence.
Details
Stated Ambition & Strategy
- Core Mission: To pursue AGI through curiosity-driven fundamental research, with open publication of models and methods as a deliberate strategy.
- Path to AGI: Emphasis on efficient architectures (Mixture-of-Experts), reinforcement-learning-based reasoning training, and squeezing maximum capability from available compute.
- Industry Impact: DeepSeek's results reframed the cost assumptions of frontier AI, prompting other labs and investors to re-examine how much spend is truly required to stay competitive.
Funding & Investors
DeepSeek was self-funded for its first years through the High-Flyer hedge fund. In 2026 it closed its first external funding round: ~$7.4 billion raised at a $52–59 billion post-money valuation, led by Tencent and CATL, with founder Liang Wenfeng making the largest individual investment (~$2.8B). A five-year lock-up with no voting rights structures the round.
Details
Key Funding Activity
- High-Flyer Self-Funding: The parent hedge fund supplied capital and a large pre-existing GPU cluster, letting DeepSeek scale without external investors.
- First External Round (2026): Closed a ~$7.4 billion (50B+ yuan) funding round — the first outside capital since founding. Post-money valuation reported at $52–59 billion (350–400B yuan).
- Key Investors: Led by Tencent (~$1.5B) and battery maker CATL (~$735M). Founder Liang Wenfeng personally invested ~20B yuan (~$2.8B), the largest individual contribution. Funds flow through a limited partnership managed by Liang Wenfeng, with a five-year lock-up and no voting rights for investors.
- Strategic Rationale: Capital raised partly to retain researchers being courted by rivals (both domestic and international), and to fund continued model development and infrastructure.
Recent Developments (2024-2026)
Released V3 (efficient MoE) and the breakout DeepSeek-R1 reasoning model (January 2025), which reshaped industry cost assumptions. Closed first external funding round in 2026: ~$7.4B at a $52–59B valuation, led by Tencent and CATL. Continued open-weight releases; latest model reported ~8 months behind top US offerings per Washington assessment.
Details
Key Announcements & Activities
- DeepSeek-V3 Release: A large, efficient Mixture-of-Experts model that demonstrated frontier-competitive performance at a fraction of typical training cost.
- DeepSeek-R1 (January 2025): An open reasoning model that matched leading proprietary systems on many benchmarks, triggering a sharp global market reaction and a wave of derivative models.
- App-Store Surge: The DeepSeek consumer app briefly topped download charts in multiple countries following the R1 launch.
- First Funding Round (2026): Closed ~$7.4 billion (50B+ yuan) at a $52–59 billion post-money valuation, led by Tencent and CATL. Founder Liang Wenfeng made the largest single investment. Five-year lock-up, no voting rights for outside investors.
- Ongoing Open Releases: Continued publishing open-weight models and technical reports, sustaining its influence over the open-model ecosystem.
Moonshot AI
Key Information
- Founded: March 2023 in Beijing, China, by Tsinghua schoolmates Yang Zhilin (CEO), Zhou Xinyu, and Wu Yuxin. [as of Jul 2026]
- Headquarters: Beijing, China.
- Valuation: ~$20B after a $2B round led by Meituan (May 2026); reportedly raising a further ~$2B at a ~$30B valuation, with a Hong Kong IPO targeted for Q4 2026 / Q1 2027. First round (Feb 2024) was led by Alibaba at ~$2.5B post-money. [as of Jul 2026]
- Flagship Models: Kimi K3 (2.8T-parameter open-weight MoE, Jul 2026) and the earlier Kimi K2 series — long-context, agentic models released with open weights.
- Main Products: The viral Kimi chatbot (web/app), the Moonshot developer API, and openly downloadable Kimi model weights.
- Official Website: moonshot.ai
- Documentation: platform.moonshot.ai
Origin & Focus
Moonshot AI was founded in 2023 by NLP researcher Yang Zhilin, a Google Brain and Meta AI veteran, around a bet on very long context windows and consumer-scale assistants. Its Kimi chatbot became one of China's most-used AI apps, and the lab has since pivoted toward frontier-scale open-weight models.
Details
Key Details
- Long-Context Origins: Kimi launched in 2023 marketing an unusually long context window (initially ~200K characters), differentiating on document and long-conversation workloads.
- Consumer Breakout: The Kimi assistant became a viral hit in China, giving Moonshot a large consumer footprint alongside its developer API.
- Open-Weight Turn: With the Kimi K2 (2025) and Kimi K3 (July 2026) releases, Moonshot moved to publishing frontier-scale open weights, competing directly with DeepSeek and Western labs.
Philosophy & Approach
Moonshot combines a consumer-product mindset with open-weight research releases, arguing that frontier capability, long context, and agentic tool-use can be delivered openly and at competitive cost. Kimi K3 is positioned as "the first open 3T-class model."
Details
Core Principles
- Scale, Openly: Kimi K3 ships as open weights at 2.8T total parameters — the largest open-weight model released to date — betting that openness plus scale wins developer mindshare.
- Efficiency at the Frontier: Sparse Mixture-of-Experts (16 of 896 experts active per token, ~1.8%), MXFP4 4-bit weights, and new attention mechanisms keep inference cost and memory manageable despite the parameter count.
- Long Context & Agents: A 1M-token context window and strong agentic/coding benchmarks target document, browsing, and tool-using workloads.
- Consumer + Developer: A dual go-to-market: the mass-market Kimi app plus an aggressively priced API and downloadable weights.
Leadership
Led by co-founder and CEO Yang Zhilin, a prominent NLP researcher (Google Brain, Meta AI; Forbes 30 Under 30 Asia 2021), alongside co-founders Zhou Xinyu and Wu Yuxin. The team is anchored by Tsinghua University talent.
Details
Key Figures
- Yang Zhilin (Kimi Yang): Co-founder and CEO; researcher behind work such as Transformer-XL and XLNet, aiming to build Moonshot in the mold of OpenAI and ByteDance.
- Zhou Xinyu & Wu Yuxin: Co-founders and Tsinghua schoolmates of Yang, part of the founding technical core.
- Strategic Backers: Alibaba, Tencent, Meituan, China Mobile, HongShan (Sequoia China), and IDG Capital among the investor base.
Key Models & Products
The Kimi line spans the K2 series and the flagship Kimi K3 — a 2.8T-parameter open-weight MoE with a 1M-token context window and native vision. Released July 16, 2026 (API); full open weights released July 27, 2026 under the Kimi K3 License. Available through the Kimi apps, Moonshot API , and open-source download.
Details
Model Line
- Kimi K3 (Jul 16, 2026): 2.8T total parameters with 16 of 896 experts active per token; adds Kimi Delta Attention (KDA) and Attention Residuals (AttnRes) for roughly 2.5× the scaling efficiency of K2. 1M-token context, native vision, MXFP4 (4-bit) weights supported natively on NVIDIA Blackwell and AMD MI400. [as of Jul 2026]
- Benchmarks (self-reported / third-party): 93.5% GPQA Diamond, 88.3% Terminal-Bench 2.1, 91.2% BrowseComp, 84.2% MCP Atlas; ranked #1 on Arena's Frontend Code eval (1,679), ahead of Claude Fable 5. Reported to trail Claude Fable 5 and GPT-5.6 Sol overall while beating Claude Opus 4.8 and GPT-5.5 on coding/agentic suites. Verify against your own eval before relying on these. [as of Jul 2026]
- Kimi K2 series (2025): The prior open-weight generation that established Kimi as a serious open-model competitor to DeepSeek.
Products & Platforms
- Kimi App: A widely used consumer chatbot in China, known for long-context document and conversation handling.
- Moonshot API: Developer access to Kimi models. K3 pricing: $0.30/M cache-hit input, $3/M input on cache miss, $15/M output tokens. [as of Jul 2026]
- Open Weights: Kimi K3's full open weights were released July 27, 2026 under the Kimi K3 License (not OSI open-source, but bespoke). Available for download and local deployment alongside the hosted API.
AGI/ASI Goals & Approach
Moonshot frames its long-term ambition around general, broadly capable AI ("moonshot") delivered to a mass audience. Its distinctive bet is that frontier capability can be reached with sparse, efficient, openly released architectures rather than closed mega-models alone.
Details
Stated Ambition & Strategy
- Core Mission: To build general-purpose AI assistants at consumer scale while pushing model capability toward the frontier.
- Path to Capability: Sparse Mixture-of-Experts, low-precision (MXFP4) training and serving, long-context attention (KDA), and agentic tool-use as the levers for capability-per-dollar.
- Open-Weight Leverage: Releasing frontier-scale weights openly to build an ecosystem and pressure closed labs on both capability and price, echoing DeepSeek's 2025 impact.
Funding & Investors
Backed by China's largest tech investors. Alibaba led the first major round (Feb 2024, ~$2.5B post-money); a ~$2B Meituan-led round (May 2026) lifted the valuation past $20B. Moonshot is reportedly raising a further ~$2B at ~$30B ahead of a planned Hong Kong IPO. ARR was reported at ~$200M as of April 2026. [as of Jul 2026]
Details
Key Funding Activity
- First Major Round (Feb 2024): Led by Alibaba at a ~$2.5B post-money valuation, establishing Moonshot as a top-tier Chinese AI startup.
- $20B Round (May 2026): Raised ~$2B in a Meituan-led round, pushing the valuation above $20B amid surging demand for open-source AI.
- Toward IPO (2026-27): Reportedly raising another ~$2B at a ~$30B valuation, with a Hong Kong IPO targeted for Q4 2026 or Q1 2027.
- Investor Base: Alibaba, Tencent, Meituan, China Mobile, HongShan (Sequoia China), and IDG Capital, among others.
Recent Developments (2024-2026)
Scaled the Kimi assistant to consumer virality, released the open-weight Kimi K2 series (2025), then launched Kimi K3 (July 16, 2026) — a 2.8T-parameter model billed as the largest open-weight release ever. Raised to a $20B+ valuation (May 2026) and is reportedly pursuing a further raise at ~$30B ahead of a Hong Kong IPO. [as of Jul 2026]
Details
Key Announcements & Activities
- Kimi K3 Launch (Jul 16, 2026): A 2.8T-parameter open-weight MoE with a 1M-token context window, claiming the top spot on Arena's Frontend Code eval and strong open-weight results on GPQA Diamond, Terminal-Bench, and BrowseComp. Full open weights slated for July 27, 2026.
- Kimi K2 Series (2025): Earlier open-weight releases that positioned Kimi alongside DeepSeek among China's leading open-model labs.
- $20B+ Valuation (May 2026): A ~$2B Meituan-led round made Moonshot one of China's most valuable AI startups.
- IPO Track (2026-27): Reportedly raising ~$2B more at ~$30B with a Hong Kong listing targeted for Q4 2026 / Q1 2027.
- Consumer Traction: Reported ARR of ~$200M (April 2026), reflecting the Kimi app's large user base.
Mistral AI
Key Information
- Founded: April 2023, by Arthur Mensch, Guillaume Lample, Timothée Lacroix. [3, 24, 30]
- Headquarters: Paris, France. [3]
- Valuation: ~€11.7 billion (~$14 billion) after a €1.7 billion Series C (September 2025, led by ASML). As of June 2026, in talks to raise €3 billion at a ~€20 billion valuation, which would nearly double its last round (discussions ongoing as of June 27). Also secured $830M in debt financing (March 2026) for datacenter buildout. Europe's most valuable AI startup.
- Flagship Models: Open-weight: Mistral 7B, Mixtral 8x7B, Mixtral 8x22B, Codestral, Mathstral, Mistral NeMo. Commercial: Mistral Large (Large 2), Mistral Small (Small 3.1), Mistral Medium, Mistral Embed, Pixtral Large (multimodal). [3, 5, 7, 25, 31]
- Main Products: La Plateforme (API for commercial models), Le Chat (conversational AI assistant, with mobile apps), open-weight models available on platforms like Hugging Face. [3, 22]
- Official Website: mistral.ai [3]
- Documentation: docs.mistral.ai
Origin & Focus
Mistral AI is a Paris-based company founded in April 2023 by former researchers from Meta AI (FAIR) and Google DeepMind. [3, 7, 24, 30] It focuses on developing open, efficient, and powerful AI models, quickly emerging as a key European player in the generative AI field. [7, 23]
Details
Key Details
- Founding Team: Arthur Mensch (CEO, previously at Google DeepMind), Guillaume Lample (Chief Scientist, previously at Meta AI), and Timothée Lacroix (CTO, previously at Meta AI). [3, 7, 24] They originally met during their studies at École Polytechnique. [3]
- Mission: To develop cutting-edge generative AI models with a strong emphasis on openness, computational efficiency, and high performance. [7, 23] They aim to be a European AI champion and democratize AI by making powerful tools accessible. [3, 7, 24]
- Rapid Emergence: Gained significant prominence and substantial funding very shortly after its inception, challenging established players with its open-weight model releases and performant commercial offerings. [24, 27]
Philosophy: Open & Efficient AI
Mistral AI strongly believes in open-weight models, typically released under permissive licenses like Apache 2.0, to foster innovation, transparency, and community building. [3, 7, 23] They focus on computational efficiency, model compactness (e.g., via Mixture-of-Experts), and providing alternatives to proprietary systems. [9, 23] Models are often available on Hugging Face . [22]
Details
Core Principles & Strategy
- Commitment to Openness: A key differentiator. Mistral AI releases many of its powerful models with open weights under licenses like Apache 2.0, allowing broad use, modification, and scrutiny by the global research and developer community. [3, 7, 21, 23] This contrasts with the more closed approach of some competitors. [9]
- Computational Efficiency: Develops models that are not only powerful but also optimized for performance, aiming for better inference speed, lower computational costs, and smaller memory footprints. This is often achieved through innovative architectures like sparse Mixture-of-Experts (MoE). [21, 23]
- Pragmatic Dual Approach: Balances its open-source contributions with optimized commercial models and API offerings (La Plateforme) for enterprise use, providing both freely accessible tools and supported enterprise-grade solutions. [22]
- European AI Leadership: Aims to build a leading AI company based in Europe, contributing to the continent's technological sovereignty and AI ecosystem, with a focus on ethical AI and privacy. [22, 24]
- Trust and Independence: Emphasizes building trustworthy AI systems and maintaining independence in its research and development roadmap.
- Democratizing AI: Seeks to make advanced AI tools more widely accessible to foster broader innovation and prevent centralization of AI power. [3, 23, 24]
Leadership
Led by co-founder and CEO Arthur Mensch. Co-founders Guillaume Lample (Chief Scientist) and Timothée Lacroix (CTO) are also key to the company's direction and technological development. [3]
Details
Key Figures
- Arthur Mensch: Co-founder and Chief Executive Officer (CEO). Formerly a researcher at Google DeepMind, with expertise in advanced AI systems and scaling laws for LLMs. [3, 7]
- Guillaume Lample: Co-founder and Chief Scientist. Formerly a researcher at Meta AI (FAIR), contributed to models like Llama. [3, 7]
- Timothée Lacroix: Co-founder and Chief Technology Officer (CTO). Formerly a researcher at Meta AI (FAIR). [3, 7]
Key Models & Products
Offers a range of open-weight models: Mistral 7B, Mixtral 8x7B, Mixtral 8x22B (MoE architecture), Codestral (code), Mathstral (math), Mistral NeMo (multilingual). [3, 23, 25, 31] Commercial models via La Plateforme API include Mistral Large (Large 2), Mistral Small (Small 3.1), Mistral Medium, Mistral Embed, and the multimodal Pixtral Large. [3, 25, 31] Key product is "Le Chat" chatbot. [3, 22]
Details
Open-Weight Models (Typically Apache 2.0 License)
- Mistral 7B: Highly efficient and performant foundational model, known for strong capabilities relative to its size (7.3 billion parameters). [3, 21, 23]
-
Mixtral Series (Sparse Mixture-of-Experts - MoE):
-
Mixtral 8x7B: Offers high performance (comparable to larger dense models) with efficient inference due to activating only a fraction of its ~47B total parameters per token. [3, 23] -
Mixtral 8x22B: A larger and more powerful open MoE model (141 billion total parameters) offering stronger performance. [3]
-
- Codestral (e.g., 22B, Mamba 7B): Specialized models for code generation, completion, and understanding. [3, 31]
- Mathstral (e.g., 7B): Specialized open-source model for mathematical reasoning and computation. [3, 31]
- Mistral NeMo (12B): Developed with NVIDIA, a fully open-source model for multilingual applications. [7, 31]
Commercial Models & Products (via La Plateforme API & Partners)
- Mistral Large (including Large 2 - 123B): Flagship commercial model series, offering top-tier reasoning capabilities, multilingual fluency (English, French, Spanish, German, Italian), and strong coding abilities. [3, 19, 25, 31]
- Mistral Small (e.g., Small 3.1 - 24B): Optimized for latency, cost-effectiveness, and efficiency, suitable for a wide range of tasks. [3, 5, 25]
- Mistral Medium (e.g., Medium 3): A mid-tier offering balancing performance and cost. [3, 5]
- Mistral Embed: State-of-the-art embedding model for tasks like semantic search and retrieval. [3, 30]
- Pixtral Large: A frontier-class multimodal model combining text and image processing. [9, 25, 31]
- Le Chat : Mistral AI's conversational AI assistant, available on the web and as mobile apps (iOS, Android), offering access to different Mistral models, web search, and image generation. [3, 22] A "Pro" version provides access to more advanced models. [3]
- La Plateforme : Mistral AI's API platform for accessing their commercial models. See docs.mistral.ai for documentation.
Platform Access & Distribution
- Open models are widely available on platforms like Hugging Face. [22]
- Commercial models are accessible via La Plateforme and through partnerships with major cloud providers like Microsoft Azure AI, Amazon Bedrock, and Google Cloud Vertex AI. [5, 25]
Approach to Advanced AI
Mistral AI focuses on building powerful and efficient foundational models. Their commitment to open-weight releases is seen as a key component for responsible and transparent AI development. While AGI is a long-term direction, the current emphasis is on tangible utility and democratizing access to advanced AI. [3, 7, 23]
Details
Perspective on AGI & Future Development
- Building Foundational Capabilities: The immediate focus is on creating highly capable and general-purpose foundational models that can serve a wide array of applications and industries.
- Efficiency as a Driver for Scale: Mistral believes that more efficient model architectures (like their use of Mixture-of-Experts) are crucial for sustainably scaling AI capabilities and making advanced models more accessible. [23]
- Openness for Safety and Broader Understanding: By releasing many models openly, Mistral AI aims to enable the global community to research their capabilities, limitations, and safety aspects. This collaborative approach is seen as vital for ensuring AI develops responsibly. [3, 7, 23, 24]
- Pragmatic and Value-Oriented Development: While the long-term trajectory of AI points towards increasingly general intelligence, Mistral's public messaging and product development prioritize delivering tangible value with existing and near-term models. Explicit AGI timelines are not a central part of their communication, focusing instead on democratizing current advanced AI. [5]
- Future Ambitions: Reports suggest plans to train models with hundreds of billions and potentially trillion parameters, aiming to achieve or surpass human-level accuracy in various NLP tasks. [5]
Funding & Partnerships
Mistral AI has rapidly raised significant funding: €105M seed (June 2023), €385M Series A (December 2023), €600M Series B (June 2024, ~$6B valuation), and €1.7B Series C (September 2025, led by ASML, ~€11.7B valuation). In March 2026, secured $830M in debt financing for datacenter expansion. As of June 2026, in talks to raise €3B at a ~€20B valuation. [24] Key investors include ASML, a16z, Lightspeed, Nvidia, and Salesforce. Strategic partnership with Microsoft includes Azure model distribution. [3, 5]
Details
Key Investment Rounds
- Seed Round (June 2023): Secured €105 million ($113 million USD), one of Europe's largest seed rounds, led by Lightspeed Venture Partners, with participation from Redpoint, Index Ventures, Xavier Niel, JCDecaux Holding, Rodolphe Saadé, Motier Ventures, La Famiglia, Headline, Exor Ventures, Sofina, First Minute Capital, and LocalGlobe. [22, 24]
- Series A (December 2023): Raised €385 million ($415 million USD), led by Andreessen Horowitz (a16z), with Lightspeed Venture Partners also significantly investing. Other participants included Salesforce, BNP Paribas, CMA CGM, General Catalyst, Elad Gil, and Nvidia. This round valued the company at approximately $2 billion. [27]
- Series B, Series C & 2026 Financing: A June 2024 Series B raised €600M at a ~$6 billion valuation. The September 2025 Series C raised €1.7 billion led by ASML, valuing Mistral at ~€11.7 billion (~$14 billion). In March 2026, Mistral secured $830M in debt financing for datacenter buildout. As of June 2026, in early-stage talks to raise €3 billion at a ~€20 billion valuation, which would nearly double its Series C valuation. ARR reached $400M+ by early 2026. [27]
Strategic Alliances & Partnerships
- Microsoft (February 2024): Announced a multi-year partnership that includes Microsoft making a €15 million investment in Mistral AI. As part of the deal, Mistral's commercial models (Mistral Large) became available on Microsoft's Azure AI platform, and the companies are collaborating on bringing models to Azure customers. [3, 5]
- Other Cloud Providers: Mistral AI models are also distributed through other major cloud platforms, including Amazon Bedrock and Google Cloud Vertex AI, expanding their enterprise reach. [5, 25]
- Nvidia: Participated in funding and collaborates on technology, including the co-development of Mistral NeMo. [7]
- Databricks, BNP Paribas: Partnerships to expand outreach and apply generative AI in specific sectors like banking. [5]
Recent Developments (2024-2025)
Released flagship Mistral Large and other commercial models (Mistral Small, Medium, Embed) via API in Feb 2024. [3] Launched open-weight Mixtral 8x22B (April 2024) and specialized models like Codestral, Mathstral, and Pixtral. [3, 25, 31] Announced strategic partnership with Microsoft (Feb 2024). [3, 5] Expanded cloud availability and launched "Le Chat" assistant with mobile apps. [3, 22] Read their news .
Details
Key Announcements & Activities
- Commercial Model Launches (Early 2024): Introduced Mistral Large, their flagship commercial model, along with Mistral Small and Mistral Embed via their "La Plateforme" API in February 2024. [3, 31]
- Open-Weight Model Releases (2024): Continued commitment to open source with releases like Mixtral 8x22B (April 2024), an open MoE model. [3] Also released specialized open models such as Codestral (for code), Mathstral (for STEM), and Codestral Mamba. [3, 31]
- Multimodal and Edge Models (Late 2024 - Early 2025): Launched Pixtral Large (multimodal text & image), and compact edge models like Ministral 3B/8B. [25, 31] Updated Mistral Small to 3.1. [5, 25]
- Strategic Partnership with Microsoft (February 2024): Announced a significant multi-year partnership including a €15 million investment from Microsoft and the availability of Mistral's models on the Azure AI platform. [3, 5]
- Cloud Platform Expansion: Models became increasingly available on other major cloud platforms like Amazon Bedrock and Google Cloud Vertex AI. [5, 25]
- "Le Chat" Conversational AI (February 2024): Launched their own AI assistant, "Le Chat," initially in beta, to provide direct access to their models. [3, 22] Mobile apps for Le Chat released in early 2025. [3]
- Continued Funding and Valuation Growth: Closed a €1.7 billion Series C (September 2025, led by ASML) at a ~€11.7 billion (~$14 billion) valuation. Secured $830M in debt financing (March 2026) for European datacenter buildout targeting 200MW by 2027 and 1GW by 2030. As of June 2026, in talks to raise €3 billion at a ~€20 billion valuation. ARR reached $400M+ by early 2026. [27]