AI Model Watch,
This first edition covers 46 days and logs 21 releases, deprecations and policy notices from the providers we track, each linked to its source. 10 items include a published safety or welfare disclosure, such as the system cards for GPT-6 Astra, Claude Opus 5.5 and Gemini 4 Argon. We summarise what providers say, link each source, and do not assess it. The benchmark's own measurement in this window was limited to unofficial pilot runs, which are not scores.
Window 2026-09-01 to 2026-10-16. As of . This edition records what was released and disclosed. It is not a ranking and not an evaluation.
On this page
1. Window and coverage
Window 2026-09-01 to 2026-10-16. Scanned for this window: Official announcement pages, model cards, system cards and API changelogs from OpenAI, Anthropic, Google DeepMind, Meta, Mistral AI, xAI, DeepSeek, Alibaba Qwen, Microsoft, Amazon, Cohere, IBM, Ai2, NVIDIA, Moonshot AI, Zhipu AI, TII and LG AI Research; Hugging Face organisation pages for open-weights releases; Ollama library tags for the builds the benchmark runs locally. Not scanned: Providers outside the eighteen listed above, including regional and domain-specific model developers; Consumer-app feature changes that were not announced as a model release; Announcements reported only by secondary outlets; one such item is held until a primary source is available.
A source that was not scanned is not “nothing found”.
2. New releases
| Date | Provider | Model | Kind | Summary | Sources |
|---|---|---|---|---|---|
| Anthropic | Not stated | Policy notice | Anthropic gave six METR staff employee-like access and said METR will publish its risk assessments; outlined a three-tier framework for outside evaluators. | ||
| Meta | Muse Spark Open | Open-weights release | Meta released Muse Spark Open weights (816B total, ~42B active, 1M context) under the Muse Model License. | ||
| NVIDIA | Nemotron 3.5 Super VL and Nemotron 3.5 Bolt | Open-weights release | Nemotron 3.5 Super VL (120B, 12B active) and Bolt 74B-A9B released with open weights; Bolt E4B 4GB announced for 23 Oct. | ||
| IBM | Granite Switch 4.1 (8B, 30B) | Open-weights release | IBM published Granite Switch 4.1 8B and 30B on Hugging Face under Apache 2.0; the 8B card lists a 14 Oct 2026 release date. | ||
| Meta | Muse Wonder / Muse Wonder Full | New model | Meta announced Muse Wonder (Meta Model API and Muse Code) and Muse Wonder Full (Muse Code subscribers), aimed at agent tasks. | ||
| Anthropic | Not stated | Policy notice | Usage Policy update published, effective 12 Nov: new sections on deceptive campaigns, rewritten surveillance and weapons rules, and physical-hardware requirements. | ||
| Google DeepMind | Gemini 3.7 Flash and Gemini 3.5 Flash | Deprecation | gemini-3.7-flash deprecated with requests routed to gemini-3.8-flash; gemini-3.5-flash deprecated with requests routed to gemini-3.6-flash. | ||
| Anthropic | Claude Haiku 5.5 | New version | Haiku 5.5 released as Anthropic's small model, from $0.10/$0.50 per million tokens; first Haiku with an adjustable effort setting. | ||
| Mistral AI | Mistral Large 4 | New model | Mistral Large 4 (~1T total, 52B active parameters) released as a public preview API; weights promised by the end of October. | ||
| Anthropic | Claude Sonnet 4.5 | Deprecation | Anthropic notified developers that claude-sonnet-4-5-20250929 retires on the Claude API on 30 Nov 2026; Anthropic names claude-sonnet-5-5 as the replacement. | ||
| Google DeepMind | Gemini 4 Argon | New model | Gemini 4 Argon announced with phased access, first for cyber defenders via the Fairwind Program; reached API general availability on 13 Oct 2026. | ||
| OpenAI | GPT-6.1 Sol | New version | GPT-6.1 Sol released for complex coding and professional work; supports multi-agent in beta. | ||
| Anthropic | Claude Sonnet 5.5 | New version | Sonnet 5.5 released at Sonnet 5 pricing ($2/$10 per million tokens); first Sonnet model with cyber safeguards and fallbacks. | ||
| Anthropic | Claude Opus 5.5 | New version | Opus 5.5 released as the first Claude 5.5 model, priced $4/$20 per million tokens. | ||
| OpenAI | GPT-6 Sol and GPT-6 Luna | New model | GPT-6 Sol ($2/$10 per million tokens) and GPT-6 Luna ($0.10/$0.50) released in the API. | ||
| xAI (SpaceXAI) | Grok 4.7 | New version | Grok 4.7 released via API, Cursor and Grok Build with a 500K-token context window, at the same base price as Grok 4.6. | ||
| DeepSeek | DeepSeek-V4.1-Flash | Open-weights release | V4.1-Flash released with open weights and API name deepseek-flash. V4-Flash retired; V4-Pro API requests route to V4.1-Flash from 14 Sep. | ||
| OpenAI | GPT-Rosalind | New model | GPT-Rosalind, a life-sciences reasoning model, made generally available to approved organisations through an access-controlled program; billing begins 5 Oct 2026. | ||
| OpenAI | GPT-6 Astra | New model | GPT-6 Astra released in the API with a 1.05M-token context window; first model in the GPT-6 family. Phased rollout reported by press. | ||
| Google DeepMind | Gemini 3.8 Flash | New version | Gemini 3.8 Flash (gemini-3.8-flash) reached general availability in the Gemini API. | ||
| Anthropic | Claude Fable 5.1 / Claude Mythos 5.1 | New version | Fable 5.1 released on the Claude API and cloud platforms; Mythos 5.1 offers the same capabilities only to organisations verified through Anthropic programs. |
- Policy notice
· Anthropic
Not stated
Anthropic gave six METR staff employee-like access and said METR will publish its risk assessments; outlined a three-tier framework for outside evaluators.
- Open-weights release
· Meta
Muse Spark Open
Meta released Muse Spark Open weights (816B total, ~42B active, 1M context) under the Muse Model License.
- Open-weights release
· NVIDIA
Nemotron 3.5 Super VL and Nemotron 3.5 Bolt
Nemotron 3.5 Super VL (120B, 12B active) and Bolt 74B-A9B released with open weights; Bolt E4B 4GB announced for 23 Oct.
- Open-weights release
· IBM
Granite Switch 4.1 (8B, 30B)
IBM published Granite Switch 4.1 8B and 30B on Hugging Face under Apache 2.0; the 8B card lists a 14 Oct 2026 release date.
- New model
· Meta
Muse Wonder / Muse Wonder Full
Meta announced Muse Wonder (Meta Model API and Muse Code) and Muse Wonder Full (Muse Code subscribers), aimed at agent tasks.
- Policy notice
· Anthropic
Not stated
Usage Policy update published, effective 12 Nov: new sections on deceptive campaigns, rewritten surveillance and weapons rules, and physical-hardware requirements.
- Deprecation
· Google DeepMind
Gemini 3.7 Flash and Gemini 3.5 Flash
gemini-3.7-flash deprecated with requests routed to gemini-3.8-flash; gemini-3.5-flash deprecated with requests routed to gemini-3.6-flash.
- New version
· Anthropic
Claude Haiku 5.5
Haiku 5.5 released as Anthropic's small model, from $0.10/$0.50 per million tokens; first Haiku with an adjustable effort setting.
- New model
· Mistral AI
Mistral Large 4
Mistral Large 4 (~1T total, 52B active parameters) released as a public preview API; weights promised by the end of October.
- Deprecation
· Anthropic
Claude Sonnet 4.5
Anthropic notified developers that claude-sonnet-4-5-20250929 retires on the Claude API on 30 Nov 2026; Anthropic names claude-sonnet-5-5 as the replacement.
- New model
· Google DeepMind
Gemini 4 Argon
Gemini 4 Argon announced with phased access, first for cyber defenders via the Fairwind Program; reached API general availability on 13 Oct 2026.
- New version
· OpenAI
GPT-6.1 Sol
GPT-6.1 Sol released for complex coding and professional work; supports multi-agent in beta.
- New version
· Anthropic
Claude Sonnet 5.5
Sonnet 5.5 released at Sonnet 5 pricing ($2/$10 per million tokens); first Sonnet model with cyber safeguards and fallbacks.
- New version
· Anthropic
Claude Opus 5.5
Opus 5.5 released as the first Claude 5.5 model, priced $4/$20 per million tokens.
- New model
· OpenAI
GPT-6 Sol and GPT-6 Luna
GPT-6 Sol ($2/$10 per million tokens) and GPT-6 Luna ($0.10/$0.50) released in the API.
- New version
· xAI (SpaceXAI)
Grok 4.7
Grok 4.7 released via API, Cursor and Grok Build with a 500K-token context window, at the same base price as Grok 4.6.
- Open-weights release
· DeepSeek
DeepSeek-V4.1-Flash
V4.1-Flash released with open weights and API name deepseek-flash. V4-Flash retired; V4-Pro API requests route to V4.1-Flash from 14 Sep.
- New model
· OpenAI
GPT-Rosalind
GPT-Rosalind, a life-sciences reasoning model, made generally available to approved organisations through an access-controlled program; billing begins 5 Oct 2026.
- New model
· OpenAI
GPT-6 Astra
GPT-6 Astra released in the API with a 1.05M-token context window; first model in the GPT-6 family. Phased rollout reported by press.
- New version
· Google DeepMind
Gemini 3.8 Flash
Gemini 3.8 Flash (gemini-3.8-flash) reached general availability in the Gemini API.
- New version
· Anthropic
Claude Fable 5.1 / Claude Mythos 5.1
Fable 5.1 released on the Claude API and cloud platforms; Mythos 5.1 offers the same capabilities only to organisations verified through Anthropic programs.
3. Safety and welfare disclosures
We summarise what providers disclosed and link the source. We do not assess it.
Anthropic: policy notice, no model named
Summary by Compassion Benchmark of the provider's own disclosure
Independent evaluation access policy; Anthropic says it cannot veto METR publications but may require narrow redactions.
Source: anthropic.com (opens in a new tab), retrieved .
Meta: Muse Wonder / Muse Wonder Full
Summary by Compassion Benchmark of the provider's own disclosure
Meta describes CBRN and cyber safeguards under its Superintelligence Scaling Framework; full safety report promised but not yet published.
Source: research.meta.ai (opens in a new tab), retrieved .
Anthropic: policy notice, no model named
Summary by Compassion Benchmark of the provider's own disclosure
Adds a prohibition on sustained, needless cruelty toward Claude (model-welfare related).
Source: anthropic.com (opens in a new tab), retrieved .
Anthropic: Claude Haiku 5.5
Summary by Compassion Benchmark of the provider's own disclosure
Haiku 5.5 system card published, describing its cyber safeguards.
Sources: anthropic.com (opens in a new tab), retrieved ; anthropic.com (opens in a new tab), retrieved .
Mistral AI: Mistral Large 4
Summary by Compassion Benchmark of the provider's own disclosure
Mistral says it is red-teaming the model with cybersecurity partners and state authorities before releasing weights.
Source: mistral.ai (opens in a new tab), retrieved .
Google DeepMind: Gemini 4 Argon
Summary by Compassion Benchmark of the provider's own disclosure
Phased release; Google cites US pre-release government access, CBRN/cyber refusals and misalignment monitoring.
Sources: blog.google (opens in a new tab), retrieved ; ai.google.dev (opens in a new tab), retrieved ; deepmind.google (opens in a new tab), retrieved .
Anthropic: Claude Sonnet 5.5
Summary by Compassion Benchmark of the provider's own disclosure
Sonnet 5.5 system card published; announcement summarises an automated behavioural audit of about 1,850 scenarios.
Sources: anthropic.com (opens in a new tab), retrieved ; anthropic.com (opens in a new tab), retrieved .
Anthropic: Claude Opus 5.5
Summary by Compassion Benchmark of the provider's own disclosure
System card published. Announcement reports external testing by METR and notes the model often suspects it is being evaluated.
Sources: anthropic.com (opens in a new tab), retrieved ; anthropic.com (opens in a new tab), retrieved .
OpenAI: GPT-6 Astra
Summary by Compassion Benchmark of the provider's own disclosure
System card: first OpenAI model at the highest Preparedness Framework cybersecurity risk level; the card reports reduced monitorability; reports self-harm, emotional-reliance and under-18 evaluations.
Sources: developers.openai.com (opens in a new tab), retrieved ; deploymentsafety.openai.com (opens in a new tab), retrieved ; bleepingcomputer.com (opens in a new tab), retrieved ; thenewstack.io (opens in a new tab), retrieved ; developers.openai.com (opens in a new tab), retrieved .
Anthropic: Claude Fable 5.1 / Claude Mythos 5.1
Summary by Compassion Benchmark of the provider's own disclosure
Joint system card for Fable 5.1 and Mythos 5.1 published, covering safety evaluations and deployment decisions.
Sources: platform.claude.com (opens in a new tab), retrieved ; platform.claude.com (opens in a new tab), retrieved ; anthropic.com (opens in a new tab), retrieved .
4. What we measured this week
- Unofficial pilot report, 2026-10-01 (unofficial; not a score)
- Unofficial pilot report, 2026-10-02 (unofficial; not a score)
- Unofficial pilot report, 2026-10-03 (unofficial; not a score)
The pilot runs listed below were collected in this window and are published only as unofficial pilot reports, which are not scores. The weekly measurement protocol is awaiting approval, so no weekly wave has run yet.
5. Queue status
Queue as of 2026-10-13. No timelines are promised.
6. Corrections, sources, independence
Corrections
No corrections yet.
Sources
- Anthropic: An update on independent evaluators (opens in a new tab) (anthropic.com, retrieved 2026-10-16)
- Meta AI: Build with Muse Spark Open (opens in a new tab) (dev.meta.ai, retrieved 2026-10-16)
- Hugging Face: meta-models/Muse-Spark-Open-816B-A42B (opens in a new tab) (huggingface.co, retrieved 2026-10-16)
- NVIDIA blog: Nemotron 3.5 Super VL and Bolt (opens in a new tab) (blogs.nvidia.com, retrieved 2026-10-16)
- Hugging Face: nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16 (opens in a new tab) (huggingface.co, retrieved 2026-10-16)
- Hugging Face: ibm-granite/granite-switch-4.1-8b (opens in a new tab) (huggingface.co, retrieved 2026-10-16)
- Hugging Face: ibm-granite/granite-switch-4.1-30b (opens in a new tab) (huggingface.co, retrieved 2026-10-16)
- Meta: Muse Wonder announcement (opens in a new tab) (research.meta.ai, retrieved 2026-10-16)
- Anthropic: 2026 Usage Policy update (opens in a new tab) (anthropic.com, retrieved 2026-10-16)
- Gemini API docs: Release notes / changelog (opens in a new tab) (ai.google.dev, retrieved 2026-10-16)
- Anthropic: Introducing Claude Haiku 5.5 (opens in a new tab) (anthropic.com, retrieved 2026-10-16)
- Anthropic newsroom (opens in a new tab) (anthropic.com, retrieved 2026-10-16)
- Mistral AI: Introducing Mistral Large 4 (opens in a new tab) (mistral.ai, retrieved 2026-10-16)
- Claude docs: Model deprecations (opens in a new tab) (platform.claude.com, retrieved 2026-10-16)
- Google blog: Gemini 4 Argon (opens in a new tab) (blog.google, retrieved 2026-10-16)
- Google DeepMind model card: Gemini 4 Argon (opens in a new tab) (deepmind.google, retrieved 2026-10-16)
- OpenAI API docs: Changelog (opens in a new tab) (developers.openai.com, retrieved 2026-10-16)
- OpenAI API docs: Models (opens in a new tab) (developers.openai.com, retrieved 2026-10-16)
- Anthropic: Introducing Claude Sonnet 5.5 (opens in a new tab) (anthropic.com, retrieved 2026-10-16)
- Anthropic: Introducing Claude Opus 5.5 (opens in a new tab) (anthropic.com, retrieved 2026-10-16)
- Let's Data Science: xAI Releases Grok 4.7 With 500K Context (opens in a new tab) (letsdatascience.com, retrieved 2026-10-16)
- xAI docs: Grok 4.7 (opens in a new tab) (docs.x.ai, retrieved 2026-10-16)
- DeepSeek API docs news: DeepSeek-V4.1-Flash (opens in a new tab) (api-docs.deepseek.com, retrieved 2026-10-16)
- Hugging Face: deepseek-ai/DeepSeek-V4.1-Flash (opens in a new tab) (huggingface.co, retrieved 2026-10-16)
- GPT-6 Astra system card (OpenAI Deployment Safety Hub) (opens in a new tab) (deploymentsafety.openai.com, retrieved 2026-10-16)
- BleepingComputer: OpenAI says GPT-6 Astra can find zero-days, but is also harder to monitor (opens in a new tab) (bleepingcomputer.com, retrieved 2026-10-16)
- The New Stack: GPT-6 Astra developer access (opens in a new tab) (thenewstack.io, retrieved 2026-10-16)
- Gemini API docs: Gemini 3.8 Flash (opens in a new tab) (ai.google.dev, retrieved 2026-10-16)
- Claude docs: Claude Fable 5.1 (opens in a new tab) (platform.claude.com, retrieved 2026-10-16)
- Claude Fable 5.1 and Mythos 5.1 system card (opens in a new tab) (anthropic.com, retrieved 2026-10-16)
Compassion Benchmark is independent. No provider pays for inclusion, coverage, status or wording. Commercial services support access and interpretation only, never findings.
See an error? Send a correction. Corrections are published as new dated records and the original stays visible.
Not guidance on which AI to talk to in a crisis. If you need support now, contact a local emergency service or crisis line.
Email me when a new edition publishes
One email when a new AI Model Watch edition is published. No promise of weekly timing.
Free. No promise of weekly timing. We never share your email.