Smallest, fastest Claude 5.5 model; $0.10 / $0.50 per million tokens.
Quick reference · last 90 days
The ten most recent qualifying releases, newest first, up to the coverage cutoff.
1T-parameter multimodal MoE in public preview; weights promised by month end.
Next Gemini frontier model, limited to vetted cyber defenders at launch.
Fastest, lowest-cost Sonnet to date at $2 / $10 per million tokens.
First Claude 5.5 model, close to Fable 5.1 on most work.
Smallest model in the new V4.1 architecture family, with visual understanding.
General Mythos-class release with 1M context and cyber safeguards.
Open multimodal MoE architectural preview; 262K native context, 1M with YaRN.
Experimental vision input added to the V4 Flash API line.
Preview moved to GA with production agent upgrades.
Master release log
Every value here comes from the release post or from the check on the verification date. Where a cell reads “not published,” the lab never stated a comparable number, and the gap was left alone rather than filled with a guess.
| Date | Model / version | Lab | Status | Modality | Context | Weights / licence | API | Input / M | Output / M | Change + official source |
|---|---|---|---|---|---|---|---|---|---|---|
| 2026-10 Two releases in the first eight days: Anthropic's Haiku 5.5 went live on the 7th, and Mistral's Large 4 opened as a public preview with weights promised by month end. | ||||||||||
| Claude Haiku 5.5 | Anthropic | GA | Not published | Not published | Closed | Yes | $0.10 (up to 100K prompt) | $0.50 (up to 100K prompt) | Smallest, fastest Claude 5.5 model; $0.50 input and $2.50 output per million above 100K prompt tokens. ANTHROPICchecked 2026-10-11 | |
| Mistral Large 4 | Mistral | PREVIEW | Text · image | Not published | Weights promised by end of October; license not named | Yes | $1.36 | $4.18 | 1T-total / 52B-active natively multimodal MoE; Mistral's largest model, in public preview. MISTRALchecked 2026-10-11 | |
| 2026-09 Anthropic's Fable 5.1 and Mythos 5.1 opened the month, DeepSeek shipped V4.1 Flash, and Anthropic followed with Opus 5.5 and Sonnet 5.5. Google's Gemini 4 Argon arrived at month's end, limited to vetted cyber defenders. | ||||||||||
| Gemini 4 Argon | LIMITED | Multimodal | Not published; 1M output | Closed | Trusted cyber defenders; paid API later | $2 intro | $10 intro | Next Gemini frontier model, rolling out via the Fairwind Program first. Price steps to $4 / $20 after the introductory period. GOOGLEchecked 2026-10-11 | ||
| Claude Sonnet 5.5 | Anthropic | GA | Not published | Not published | Closed | Yes | $2 | $10 | Fastest, lowest-cost Sonnet to date; upgrades Sonnet 5 and works beside Opus 5.5. ANTHROPICchecked 2026-10-11 | |
| Claude Opus 5.5 | Anthropic | GA | Not published | Not published | Closed | Yes | $4 | $20 | First Claude 5.5 model; close to Fable 5.1 on most work at 40% lower run cost than Opus 5. ANTHROPICchecked 2026-10-11 | |
| DeepSeek-V4.1-Flash | DeepSeek | RELEASE | Text · image | Not published | Open weights linked; terms not checked | Yes | Peak/off-peak schedule | Peak/off-peak schedule | Smallest model in the new V4.1 architecture family; V4 Pro requests rerouted to it from 2026-09-14. DEEPSEEKchecked 2026-10-11 | |
| Claude Fable 5.1 | Anthropic | GA | Not published | 1M; 128K output | Closed | Yes | $10 | $50 | General-use Mythos-class release; cache reads $0.25. Cyber safeguards block exploit development but allow vulnerability finding. ANTHROPICchecked 2026-10-11 | |
| Claude Mythos 5.1 | Anthropic | CONTROLLED | Not published | 1M; 128K output | Closed | Restricted to verified organizations | $10 | $50 | Sibling with more permissive safeguards for vetted users, same pricing as Fable 5.1 per platform docs. ANTHROPICchecked 2026-10-11 | |
| 2026-08 Two labs, five rows, no Western frontier launch at all. Qwen showed an early look at the Qwen4 architecture and its first Max-class model with open weights; DeepSeek took V4 Pro out of preview and bolted experimental vision onto V4 Flash. | ||||||||||
| Qwen3.8-Flash-Next | Qwen | RELEASE | Text · image | 262K native; 1M YaRN | Open weights; Apache 2.0 | Coming soon at post | $0.16 hosted | $0.47 hosted | Architectural preview for Qwen4 with 125B main + 51B n-gram parameters, 6B active. QWENchecked 2026-09-01 | |
| Qwen3.8-Flash-Next-Base | Qwen | CHECKPOINT | Text · image | 262K native; 1M YaRN | Open weights; Apache 2.0 | No managed base API | Self-host | Self-host | Base checkpoint released beside the post-trained model for independent architecture evaluation. QWENchecked 2026-09-01 | |
| DeepSeek-V4-Flash-Vision-Exp | DeepSeek | EXPERIMENT | Text · image | 1M family context | Closed API | Yes | V4 Flash rate | V4 Flash rate | Added mixed image/text input; images billed as up to 384 tokens each. DEEPSEEKchecked 2026-09-01 | |
| DeepSeek-V4-Pro GA | DeepSeek | GA | Text | 1M | Open weights | Yes | Peak/off-peak schedule | Peak/off-peak schedule | Promoted V4 Pro from preview with flexible reasoning effort and Responses API support. DEEPSEEKchecked 2026-09-01 | |
| Qwen3.8-Max | Qwen | RELEASE | Text · image | 1M | Open weights announced | Yes | See QwenCloud | See QwenCloud | 2.4T-total / 95B-active Max-class model for coding and cowork. QWENchecked 2026-09-01 | |
| 2026-07 One flagship, Claude Opus 5, and then specialists: a third-generation Qwen image model, an open proof-coding checkpoint from Mistral, and Mistral’s first model for moving a robot through a space. | ||||||||||
| Claude Opus 5 | Anthropic | GA | Text · image | 1M | Closed | Yes | $5 | $25 | Daily-use Opus release near Fable capability at half Fable input price. ANTHROPICchecked 2026-09-01 | |
| Qwen-Image-3.0 | Qwen | RELEASE | Image generation | 4.5K input prompt | Closed at announcement | Qwen Studio | Not token-comparable | Not token-comparable | Third-generation image model with dense-layout and 12-language text rendering. QWENchecked 2026-09-01 | |
| Robostral Navigate | Mistral | RELEASE | Vision · action | Not published | See model card | Research release | Not applicable | Not applicable | Mistral’s first model built for embodied navigation. MISTRALchecked 2026-09-01 | |
| Leanstral 1.5 | Mistral | RELEASE | Text · code | Not published | Open source | Weights | Self-host | Self-host | New proof-oriented coding foundation checkpoint. MISTRALchecked 2026-09-01 | |
| 2026-06 The heaviest month in the log by a wide margin. Anthropic opened its Fable and Mythos tier on the 9th and pushed frontier work down into Sonnet three weeks later, OpenAI previewed three GPT-5.6 tiers in a single post, and Google spent the month on Gemma checkpoints before closing with a text model that generates by diffusion instead of token by token. | ||||||||||
| Claude Sonnet 5 | Anthropic | GA | Text · image | 1M | Closed | Yes | $2 at launch | $10 at launch | Frontier agentic performance moved into the balanced Sonnet tier. ANTHROPICchecked 2026-09-01 | |
| GPT-5.6 Sol | OpenAI | PREVIEW | Text · image | 1.05M | Closed | Limited | $5 | $30 | Flagship of the preview series; introduced max and multi-agent ultra modes. Superseded 2026-09-22 by GPT-6 Sol at $2 / $10. OPENAIchecked 2026-09-01 | |
| GPT-5.6 Terra | OpenAI | PREVIEW | Text · image | 1.05M | Closed | Limited | $2 | $12 | Balanced series member, announced as competitive with GPT-5.5 at half price. No GPT-6 Terra; GPT-6 Sol took the workhorse tier on 2026-09-22. OPENAIchecked 2026-09-01 | |
| GPT-5.6 Luna | OpenAI | PREVIEW | Text · image | 1.05M | Closed | Limited | $0.20 | $1.20 | Fast, lowest-cost member of the GPT-5.6 preview family. Superseded 2026-09-22 by GPT-6 Luna at $0.10 / $0.50. OPENAIchecked 2026-09-01 | |
| Mistral OCR 4 | Mistral | GA | Document vision | Page-based | Closed; self-host option | Yes | $4 / 1K pages | Structured output | Added bounding boxes, block classes and inline confidence across 170 languages. MISTRALchecked 2026-09-01 | |
| DiffusionGemma 26B MoE | EXPERIMENT | Text | Not published | Open; Apache 2.0 | Weights | Self-host | Self-host | Experimental block-diffusion text model, reported up to 4× faster on dedicated GPUs. GOOGLEchecked 2026-09-01 | ||
| Claude Fable 5 | Anthropic | GA | Text · image | 1M | Closed | Yes | $10 | $50 | First generally released Mythos-class capability with cyber request routing safeguards. ANTHROPICchecked 2026-09-01 | |
| Claude Mythos 5 | Anthropic | CONTROLLED | Text · image | 1M | Closed | Restricted access | $10 | $50 | More cyber-capable sibling released under tighter access; source records later suspension/redeployment. ANTHROPICchecked 2026-09-01 | |
| Gemma 4 E2B QAT | CHECKPOINT | Text · image · audio | 128K | Open; Apache 2.0 | Weights | Self-host | Self-host | Q4_0 and mobile quantization-aware checkpoints; mobile format targets a 1GB footprint. GOOGLEchecked 2026-09-01 | ||
| Gemma 4 E4B QAT | CHECKPOINT | Text · image · audio | 128K | Open; Apache 2.0 | Weights | Self-host | Self-host | Quantization-aware edge checkpoint released for Q4_0 and mobile runtimes. GOOGLEchecked 2026-09-01 | ||
| Gemma 4 12B QAT | CHECKPOINT | Text · image · audio | 256K | Open; Apache 2.0 | Weights | Self-host | Self-host | Quantization-aware laptop checkpoint followed the 12B unified model by two days. GOOGLEchecked 2026-09-01 | ||
| Gemma 4 26B MoE QAT | CHECKPOINT | Text · image | 256K | Open; Apache 2.0 | Weights | Self-host | Self-host | QAT checkpoint reduced deployment memory while preserving the 26B MoE family behavior. GOOGLEchecked 2026-09-01 | ||
| Gemma 4 31B QAT | CHECKPOINT | Text · image | 256K | Open; Apache 2.0 | Weights | Self-host | Self-host | QAT checkpoint for the dense 31B family release. GOOGLEchecked 2026-09-01 | ||
| Gemma 4 12B | RELEASE | Text · image · audio | 256K | Open; Apache 2.0 | AI Studio + weights | See platform | See platform | Unified encoder-free mid-size model brought native audio input to laptops. GOOGLEchecked 2026-09-01 | ||
| 2026-05 Google I/O carried May with Gemini Omni and the first Gemini 3.5 model. Around it: four Gemma speculative-decoding drafters on one day, Opus 4.8, Mistral Medium 3.5, and Qwen’s generalist robot policy. | ||||||||||
| Qwen-VLA | Qwen | RESEARCH | Vision · language · action | Not published | Research release | No general API | Not applicable | Not applicable | Generalist embodied policy spanning manipulation and navigation. QWENchecked 2026-09-01 | |
| Claude Opus 4.8 | Anthropic | GA | Text · image | 1M | Closed | Yes | $5 | $25 | More reliable agentic collaborator; added user effort control and cheaper fast mode. ANTHROPICchecked 2026-09-01 | |
| Mistral Medium 3.5 | Mistral | PREVIEW | Text · image | 256K | Open weights; modified MIT | Yes | $1.50 | $7.50 | 128B dense upgrade for long-horizon coding and productivity work. MISTRALchecked 2026-09-01 | |
| Gemini Omni | RELEASE | Any input · video-first output | Not published | Closed | Staged access | Not published | Not published | Native multimodal creation and conversational editing model introduced at I/O. GOOGLEchecked 2026-09-01 | ||
| Gemini 3.5 Flash | RELEASE | Text · image · audio · video | 1M | Closed | Yes | See Gemini API | See Gemini API | First Gemini 3.5 family model, tuned for agentic workflows and action. GOOGLEchecked 2026-09-01 | ||
| Gemma 4 E2B MTP drafter | CHECKPOINT | Draft tokens | 128K target | Open; Apache 2.0 | Weights | Self-host | Self-host | Speculative-decoding drafter checkpoint for lower-latency E2B inference. GOOGLEchecked 2026-09-01 | ||
| Gemma 4 E4B MTP drafter | CHECKPOINT | Draft tokens | 128K target | Open; Apache 2.0 | Weights | Self-host | Self-host | Speculative-decoding drafter for the E4B edge model. GOOGLEchecked 2026-09-01 | ||
| Gemma 4 26B MTP drafter | CHECKPOINT | Draft tokens | 256K target | Open; Apache 2.0 | Weights | Self-host | Self-host | MTP drafter reduced sequential decoding latency for the MoE model. GOOGLEchecked 2026-09-01 | ||
| Gemma 4 31B MTP drafter | CHECKPOINT | Draft tokens | 256K target | Open; Apache 2.0 | Weights | Self-host | Self-host | MTP drafter released for the dense 31B checkpoint. GOOGLEchecked 2026-09-01 | ||
| GPT-5.5 Instant | OpenAI | CHATGPT | Text · image | Product-managed | Closed | No distinct API model | Plan-based | Plan-based | Updated ChatGPT default model for clearer, more personalized everyday responses. OPENAIchecked 2026-09-01 | |
| 2026-04 Two days in late April carried GPT-5.5, its Pro sibling, and both DeepSeek V4 previews. Earlier in the month Google opened the Gemma 4 family with four sizes at once, and Meta’s new Superintelligence Labs shipped its first Muse model. | ||||||||||
| GPT-5.5 Pro | OpenAI | GA API | Text · image | 1M class | Closed | Yes | $5 | $30 | Maximum-performance sibling entered the API in the release post’s April 24 update. OPENAIchecked 2026-09-01 | |
| DeepSeek-V4-Pro Preview | DeepSeek | PREVIEW | Text | 1M | Open weights | Yes | See API pricing | See API pricing | 1.6T-total / 49B-active flagship with thinking and non-thinking modes. DEEPSEEKchecked 2026-09-01 | |
| DeepSeek-V4-Flash Preview | DeepSeek | PREVIEW | Text | 1M | Open weights | Yes | See API pricing | See API pricing | 284B-total / 13B-active efficient sibling with dual reasoning modes. DEEPSEEKchecked 2026-09-01 | |
| GPT-5.5 | OpenAI | RELEASE | Text · image | 1M class | Closed | API 2026-04-24 | $5 | $30 | Step up in agentic coding, knowledge work, computer use and scientific workflows. OPENAIchecked 2026-09-01 | |
| Qwen3.6-Max-Preview | Qwen | PREVIEW | Text | Not published | Closed hosted model | Coming soon at post | Not published | Not published | Preview improved agentic coding, knowledge and instruction following over 3.6 Plus. QWENchecked 2026-09-01 | |
| Claude Opus 4.7 | Anthropic | GA | Text · image | 1M | Closed | Yes | $5 | $25 | Advanced software-engineering and vision upgrade with new cyber safeguards. ANTHROPICchecked 2026-09-01 | |
| Muse Spark | Meta | RELEASE | Native multimodal | Not published | Closed | Private preview | Not published | Not published | First Meta Superintelligence Labs Muse model, with tool use and visual reasoning. METAchecked 2026-09-01 | |
| Gemma 4 E2B | RELEASE | Text · image · audio | 128K | Open; Apache 2.0 | Weights / AI Edge | Self-host | Self-host | Effective-2B edge model for on-device multimodal workloads. GOOGLEchecked 2026-09-01 | ||
| Gemma 4 E4B | RELEASE | Text · image · audio | 128K | Open; Apache 2.0 | Weights / AI Edge | Self-host | Self-host | Effective-4B edge member with native audio input. GOOGLEchecked 2026-09-01 | ||
| Gemma 4 26B MoE | RELEASE | Text · image | 256K | Open; Apache 2.0 | AI Studio + weights | See platform | See platform | 26B mixture-of-experts model activating 3.8B parameters per token. GOOGLEchecked 2026-09-01 | ||
| Gemma 4 31B Dense | RELEASE | Text · image | 256K | Open; Apache 2.0 | AI Studio + weights | See platform | See platform | Dense quality-oriented workstation model in the initial Gemma 4 family. GOOGLEchecked 2026-09-01 | ||
| 2026-03 Five rows, three of them Mistral’s: a first text-to-speech model, the hybrid Small 4, and the open Leanstral proof-coding foundation. OpenAI’s GPT-5.4 pair added native computer use and tool search. | ||||||||||
| Voxtral TTS 4B | Mistral | GA | Text to speech | Not applicable | CC BY-NC 4.0 weights | Yes | $16 / M chars | Audio output | First Mistral TTS model; nine languages and streaming voice generation. MISTRALchecked 2026-09-01 | |
| Mistral Small 4 | Mistral | GA | Text · image | 256K | Open; Apache 2.0 | Yes | $0.15 | $0.60 | 119B-total / 6.5B-active hybrid unifying instruct, reasoning and coding. MISTRALchecked 2026-09-01 | |
| Leanstral | Mistral | RELEASE | Text · code | Not published | Open source | Weights | Self-host | Self-host | Open foundation model for proof-oriented trustworthy coding. MISTRALchecked 2026-09-01 | |
| GPT-5.4 | OpenAI | GA | Text · image · computer use | 1M | Closed | Yes | $2.50 | $15 | General professional model added native computer use and tool search. OPENAIchecked 2026-09-01 | |
| GPT-5.4 Pro | OpenAI | GA | Text · image · computer use | 1M | Closed | Yes | $30 | $180 | Maximum-performance reasoning sibling released in ChatGPT and API. OPENAIchecked 2026-09-01 | |
| 2026-02 The year’s first crowded month. OpenAI led with coding, including a Codex preview served above 1,000 tokens per second on Cerebras hardware. Anthropic shipped Opus 4.6 and Sonnet 4.6, Qwen opened its 3.5 line, and Mistral rebuilt its transcription stack. | ||||||||||
| Claude Sonnet 4.6 | Anthropic | GA | Text · image | 1M beta | Closed | Yes | $3 | $15 | Sonnet-tier upgrade for coding, agents and professional work. ANTHROPICchecked 2026-09-01 | |
| Qwen3.5-397B-A17B | Qwen | RELEASE | Text · image | 262K native | Open weights; Apache 2.0 | Weights | Self-host | Self-host | First open Qwen3.5 model; hybrid attention MoE with 17B active parameters. QWENchecked 2026-09-01 | |
| Qwen3.5-Plus | Qwen | HOSTED | Text · image | 1M | Closed hosted model | Yes | See Model Studio | See Model Studio | Hosted family member with built-in tools and adaptive tool use. QWENchecked 2026-09-01 | |
| GPT-5.3-Codex-Spark | OpenAI | RESEARCH | Text | 128K | Closed | Design partners | Not listed | Not listed | Ultra-fast coding preview served above 1,000 tokens/s on Cerebras. OPENAIchecked 2026-09-01 | |
| Qwen-Image-2.0 | Qwen | RELEASE | Image generation · editing | 1K prompt | Closed at launch | Qwen Studio | Not token-comparable | Image output | Unified 2K generation/editing model with long-form typography support. QWENchecked 2026-09-01 | |
| GPT-5.3-Codex | OpenAI | RELEASE | Text · image | Codex-managed | Closed | Not at launch | Plan-based | Plan-based | Unified Codex and GPT training; 25% faster than GPT-5.2-Codex. OPENAIchecked 2026-09-01 | |
| Claude Opus 4.6 | Anthropic | GA | Text · image | 1M beta | Closed | Yes | $5 | $25 | Longer agentic work, adaptive thinking, effort controls and compaction. ANTHROPICchecked 2026-09-01 | |
| Voxtral Mini Transcribe V2 | Mistral | GA | Speech to text | Audio stream | Closed | Yes | Audio-priced | Transcript | Batch transcription with diarization, biasing and word timestamps in 13 languages. MISTRALchecked 2026-09-01 | |
| Voxtral Realtime | Mistral | GA | Speech to text | Live audio | Open; Apache 2.0 | Yes | Audio-priced | Transcript | Live transcription with configurable latency down to sub-200ms. MISTRALchecked 2026-09-01 | |
| Qwen3-Coder-Next | Qwen | RELEASE | Text · code | Not published | Open weights | Weights | Self-host | Self-host | Agentic coding model based on Qwen3-Next-80B-A3B-Base. QWENchecked 2026-09-01 | |
| 2026-01 One row: a DeepMind world model with no general API. Plenty of product news shipped in January, but none of it was a dated model release, so none of it is here. | ||||||||||
| Project Genie | Google DeepMind | EXPERIMENT | Prompt · image → interactive world | World-state session | Closed | No general API | Not applicable | Not applicable | Experimental general-purpose world model for interactive generated environments. DEEPMINDchecked 2026-09-01 | |
What “release” means here
Dates only compare if you resolve them the same way every time. This page takes the earliest date the lab itself says the named model became usable by whoever the lab named.
DATE RESOLUTION PROTOCOL
- GA / release: use the announcement date when access begins that day.
- Preview: include only when named users can actually access it; stamp PREVIEW or LIMITED.
- API after app: keep the first release row and state the later API date in availability.
- Checkpoint variants: separate rows only when the lab explicitly releases separate artifacts.
- Redeployment: availability changes belong in status notes, not as a second model launch.
AMBIGUITIES TO PRESERVE
Preview versus GA
A preview does not quietly turn into GA here. DeepSeek V4 Pro gets a preview row on April 24 and a GA row on August 13, and the near-four-month gap between them is the part worth knowing.
Staged and region-limited access
The date is the first documented access date, and the restriction goes in the status and API cells. A date here never means you personally could get in that day.
Silent model swaps
When a lab changes model behavior behind the same name, it stays out until the lab gives that change a dated release identity of its own. Product release notes alone do not earn a row.
Version lineages
Square = general release, circle = preview or controlled access, filled = open weights. Retirement dates appear where a lab published one; a missing date means the lab had not published one by the verification cutoff.
OpenAI GPT / Codex
5.3 Codex02-05
Codex Spark02-12
5.4 / Pro03-05
5.5 / Pro04-23/24
5.6 family06-26
Anthropic Claude
Opus 4.602-05
Sonnet 4.602-20
Opus 4.704-16
Opus 4.805-28
Fable/Mythos 506-09
Sonnet 506-30
Opus 507-24
Fable 5.109-01
Opus 5.509-22
Sonnet 5.509-28
Haiku 5.510-07
Google Gemma
Gemma 4 ×404-02
MTP ×405-05
12B06-03
QAT ×506-05
Diffusion06-10
Qwen core
Coder Next02-02
3.5 397B02-15
3.6 Max04-18
3.8 Max08-02
3.8 Flash Next08-26
DeepSeek V4
V4 Pro04-24 preview
V4 Flash04-24 preview
V4 Pro08-13 GA
Flash Vision08-21 exp
V4.1 Flash09-10
Mistral Small / Medium
Small 403-16
Medium 3.505-22
Large 410-06 preview
Mistral Voxtral
Mini Transcribe V202-04
Realtime02-04
TTS03-23
Qwen Image
Image 2.002-10
Image 3.007-21
What changed, in aggregate
Everything below is counted off the 68 rows above, so it describes this register rather than the whole model market. Separately released checkpoints each count as one.
Releases per quarter
Q2 ran away with the year because Gemma’s checkpoint program and several frontier families landed inside the same twelve weeks.
Open-weight share
28 of 60 rows ship weights you can download under stated terms. Open here means you can run it yourself and the licence is named. Read that licence before assuming what else you are allowed to do with it.
OPEN-WEIGHT 47%CLOSED / CONTROLLED 53%
Context ceiling steps
The largest comparable text context at each step: 128K on edge models, 256K on the open laptop checkpoints, then a cluster of 1M-class API families from April onward.
Frontier input-price movement
Indexed to GPT-5.4’s $2.50/M input as 100. One vendor’s headline prices only. Previews and Pro tiers stay in the line on purpose, since averaging them away would hide the swings worth seeing.
Common mistakes / anti-patterns
How release histories become wrong even when every link is real.
ANNOUNCED ≠ USABLE
Record a preview only when the announcement names an audience with access. A roadmap promise does not get a row.
FAMILY ≠ ENDPOINT
One announcement post can name four or five artifacts. Write down the exact model version and the channel you reach it through, because Gemma 4 on its own does not say which of the four sizes you meant.
CURRENT PRICE ≠ LAUNCH PRICE
This table holds the headline price as checked on the verification date. Open the official pricing page before you budget, because rates, batch discounts and cache pricing all drift.
Coverage and method
This log survives weekly upkeep because it refuses most of what crosses it. Narrow lab list, hard inclusion rule, blanks left blank, and a cutoff date printed where you can see it.
Tracked labs
- OpenAI, Anthropic, Google DeepMind / Google, Meta
- Mistral, DeepSeek, Alibaba Qwen
- Other frontier or near-frontier trainers enter when a qualifying primary announcement is found
A log built from announcements over-counts the labs that announce. Quiet model swaps and labs that never post release notes are under-counted here, and nothing short of probing endpoints would fix that.
Update protocol
- Check official lab news, model cards, API release notes and pricing docs weekly.
- Cap a routine update at 10 new rows; on anomalous dates or conflicting sources, change nothing and report.
- Never move an existing date without keeping the old source beside the new one and comparing them.
- Report omissions via GitHub Issues with the lab’s announcement URL.