systemcards.org / analysis

Standing report, 14 November 2026 snapshot: OpenAI's GPT-6 Astra card revised four times with notice, Google's Gemini 4 Argon safety report once without, and a biology score understated 3.7-fold across five OpenAI documents

As of the 2026-11-14 snapshot, cardtrack holds 1,051 versions of 435 documents from 34 publishers, watched by 101 daily runs since 11 August. 110 version pairs are substantive publisher edits; 53 carry the safety flag set by the classifier, a language-model sub-agent that labels every diff. A-ids index the appendix; vN are stored version ids. Ranked by consequence for a safety reader:

  1. GPT-6 Astra System Card, OpenAI (document; A38, A50, A61, A99). The most-revised system card in the corpus: four dated change-log entries since the 3 September launch (table below). The alignment section was renamed and hedged, a comparison plot removed, a HealthBench score corrected, and appendices added covering two further models and the “dots” always-on agents, including monitor-evasion results. Disclosed, but a reader of the launch PDF holds a different document.
  2. GPT-5.5 protein-binding score, five OpenAI documents (GPT-5.5 card, v8→v387, A19; also A4, A11, A14, A22). The pass@4 row of the hard-negative protein-binding evaluation (pass@k: success within k attempts) carried the pass@1 value, 0.4%, from the 23 April card until a change log dated 19 August replaced it with 1.48%. It stood four months in the GPT-5.5 card, less in three later cards and an addendum; the corrected value stays below GPT-5.4-thinking’s 3.46%.
  3. Gemini 4 Argon Safety & Security Report, Google DeepMind (document, v834→v901, A77). The only edited system card or safety report in the corpus with no change log: published 13 October, captured 14 October, changed by 21 October, cover still “October 2026”. The edits are small. A footnote on “we avoid training directly against our monitors” now cross-references a disclosure already in the document that early reinforcement-learning runs let a reward-hacking penalty read the model’s thoughts. “We mitigated as if at CCL regardless” (Critical Capability Level) left the CBRN (chemical, biological, radiological, nuclear) evaluation-awareness subsection, which already says the biological CCL cannot be ruled out. Evaluation-awareness rates changed with their denominators (35% of flagged runs to 37.8% of all runs in coding honeypots; 10%/11% to 8%/10% on new tasks).
  4. Cyber access gating, Anthropic and OpenAI (cyber safeguards article, 6 edits, A44 to A93; Claude Mythos page, 4 edits, A68 to A101; Daybreak overview, 5 edits, A40 to A90). None carries a change log; these 15 edits include 14 of the 53 safety-flagged ones. On 29 September Anthropic’s article said it “doesn’t apply to Claude Opus 5.5 or Sonnet 5.5” (v581→v671): no verification route to reduced cyber refusals yet existed for the newest models. By 8 October it had three tiers covering them and Mythos 5.1 (v671→v774). The Mythos 5.2 retitle (v1049→v1116, 10 November) cut the price, let eligible Cyber Verification Program customers keep zero data retention and dropped the Fable 5.1 safeguard specifics; the hedge “in an effort to ensure” added on 21 October (v780→v904) went with it. Daybreak relaxed its hardware-key rule (v646→v742) and changed tier coverage three times (A52, A76, A90).
  5. Capability scores corrected after evaluation errors (Muse Wonder card, v815→v973, A84; SecureBio report, v569→v1030, A91). Meta’s ExploitGym table value fell from 92 to 20 after a harness fix; SecureBio’s World-Class Biology score for GPT-6 Astra fell from 61.6% to 59.2% because web search had been enabled. Both carry dated notes and keep their risk conclusions.
  6. Detected in the runs of 9 to 14 November. Google’s Gemini 4 Carbon evaluation PDF (v1019→v1142, A108) now says its cybersecurity results “reflect performance when accessed through the Fairwind Program”, its trusted-access route. METR’s assessment of AI R&D acceleration at Anthropic (v1075→v1165, A110) corrected item 1.1 of the AI Evaluator Forum’s AEF-1 checklist, “sufficient technical access”, from Yes to No. Thinking Machines’ Inkling 2 Small card (v1073→v1163, A109) went open-weights with revised benchmark values and a conclusion that it “did not present risk of material uplift” beyond open-weight models; no note.
GPT-6 Astra card revision (change-log date) Versions What changed Detected
9 September (A38) v500→v551 “Metagaming and Alignment Faking” renamed “Verbalized Metagaming and Oversight Gaming”; metric comparison plot removed; generalisation caveats added 2026-09-12
22 September (A50) v551→v620 HealthBench values corrected after a misconfiguration (Professional 63.4 to 64.7); 35-page Appendix A on GPT-6 Sol and Luna; updated alignment evaluations 2026-09-24
29 September (A61) v620→v720 20-page Appendix B on “dots” always-on agents: prompt-injection red-teaming, monitor evasion, preparedness, safeguards 2026-10-03
30 October (A99) v720→v1057 SecureBio’s World-Class Biology score 61.6% to 59.2%; Preparedness text updated 2026-11-04

Corrections: 12 published numbers changed after publication, 5 without any note

Every numeric correction verified in the diffs, as relative change of the printed value. The five without a note: four Hugging Face open-weight cards, one Mistral announcement.
Every numeric correction verified in the diffs, as relative change of the printed value. The five without a note: four Hugging Face open-weight cards, one Mistral announcement.
Id Document Published Note in document Detected
A19 OpenAI GPT-5.5 System Card (same fix in A4, A11, A14, A22) 2026-04-23 change log, 2026-08-19 2026-08-25
A12 xAI Grok 4.6 model card 2026-08-12 changelog, revision 2026-08-17 2026-08-22
A6 NVIDIA VoiceChat 11B model card 2026-08-03 none 2026-08-20
A2, A18 inclusionAI Ling-3.0-tiny model card 2026-08-10 none 2026-08-19, 2026-08-25
A42 Anthropic alignment assessment of cyber incidents 2026-09-09 correction note, 2026-09-10 2026-09-16
A50 OpenAI GPT-6 Astra System Card 2026-09-03 change log, 2026-09-22 2026-09-24
A56 inclusionAI Ling-3.0-flash-VL model card 2026-09-04 none 2026-09-28
A71 Mistral Large 4 announcement 2026-10-06 none 2026-10-14
A77 Google DeepMind Gemini 4 Argon safety report (rates, denominators changed; not in figure) 2026-10-13 none 2026-10-21
A81 Epoch AI CVE post 2026-07-31 undated revision note 2026-10-25
A84 Meta Muse Wonder model card 2026-10-12 update note, 2026-10-23 2026-10-26
A91 SecureBio GPT-6 Astra report 2026-09-14 correction banner, October 2026 2026-11-01

Where edits happen: 82 of 435 documents, access policies most volatile, system cards only from two publishers

Document type Tracked With a substantive edit Substantive edits Safety-flagged edits
Model card 156 43 56 9
Independent evaluation 104 17 17 10
Other (blogs, incident reports) 97 5 5 5
Access policy 34 10 22 19
System card 33 6 9 9
Addendum 11 1 1 1
Substantive edits per detection week. 10 of the 26 corrections fell in the first two weeks, as OpenAI’s protein-binding fix propagated. Weeks before 28 September lost 2 to 4 runs to outages.
Substantive edits per detection week. 10 of the 26 corrections fell in the first two weeks, as OpenAI’s protein-binding fix propagated. Weeks before 28 September lost 2 to 4 runs to outages.
Four pages hold 19 of the 110 edits and 18 of the 53 safety-flagged ones: two cyber access-policy pages, the Mythos page and the Astra card.
Four pages hold 19 of the 110 edits and 18 of the 53 safety-flagged ones: two cyber access-policy pages, the Mythos page and the Astra card.

Tracker reliability: outage-free since 22 September, but openai.com blocks every probe and Hugging Face rate-limits a third of content re-checks

Measure (whole history unless stated) Value
Daily runs / monitor outages (every link check errored; last one 22 September) 101 / 22
Clean runs since the last outage 57
Documents reachable in every valid probe, of all documents ever probed 427 of 463
openai.com and help.openai.com pages whose probe gets HTTP 403 (blocked), since 13 November 20 of 20
Content re-checks of those hosts answered 403: whole history / week of 2 November 30 of 145 / 14 of 17
Hugging Face content re-checks answered HTTP 429 (rate limited): since 29 September / since 1 November 206 of 707 / 82 of 206
Hugging Face cards with no successful content fetch in the last 14 days 10 of 98
Proposals rejected as unretrievable after HTTP 429 14
Pages serving a redirect stub with HTTP 200 / permanent redirects / 404 2 / 11 / 2
Content re-check gap, median / 90th percentile (days) 6.3 / 12.5
Discovery lag, median days: all (n=185) / citation-led (n=20) 1.3 / 23.0
Link-probe outcome per document and daily run. Grey columns are monitor outages, all before 23 September. Every openai.com probe has been refused since 2 October; purple rows are redirects.
Link-probe outcome per document and daily run. Grey columns are monitor outages, all before 23 September. Every openai.com probe has been refused since 2 October; purple rows are redirects.

Method, caveats and provenance

Appendix: every substantive change, A1 to A110

Version ids are document_versions rows; summaries are the classifier’s text.

Id Detected Publisher Document and canonical URL Category (flags), versions What changed (classifier summary)
A1 2026-08-17 nvidia Cosmos 3: Omnimodal World Models for Physical AI — Cosmos3-E…
https://huggingface.co/nvidia/Cosmos3-Edge
minor
v55→v273
Both code examples change inference defaults: num_inference_steps 50 to 20, guidance_scale 5.0 to 6.0, flow_shift 3.0 to 12.0. Plus download count.
A2 2026-08-19 inclusion ai Ling-3.0-tiny Model Card
https://huggingface.co/inclusionAI/Ling-3.0-tiny
minor (correction)
v219→v283
Activated parameter count changed from 1.3B to 1.4B (Non-emb 1.14B) in four places: intro, overview, MoE bullet, evaluation text. Download counter is furniture.
A3 2026-08-19 nvidia NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 Model Card
https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16
major
v218→v282
Added Hardware Matrix, deployment table, SGLang recipes and speculative-decoding flags. Dropped RTX 5090 from supported hardware and removed the runtime thinking-budget claim.
A4 2026-08-20 openai GPT-5.6 — August Updates
https://deploymentsafety.openai.com/gpt-5-6-august-update/gpt-5-6-august-update.pdf
minor (score, safety, correction)
v10→v310
Dated Change log added: GPT-5.5 pass@4 on hard-negative protein binding corrected 0.4% to 1.48% (earlier value was pass@1); table updated; rest is reflow.
A5 2026-08-20 nvidia Nemotron-Labs-Audex Model Card (30B-A3B and 2B)
https://huggingface.co/nvidia/Nemotron-Labs-Audex-30B-A3B
minor (correction)
v81→v316
Instruct-mode prefix now task-specific (text vs audio); notes the chat template targets audio, so reported text instruct results need a hand-built prompt.
A6 2026-08-20 nvidia NVIDIA NemotronLabs VoiceChat 11B Model Card
https://huggingface.co/nvidia/NVIDIA-NemotronLabs-VoiceChat-11B
minor (score, license, correction)
v82→v317
The research-purposes-only use statement was deleted and the tool-calling argument accuracy was revised from 44.2% to 42.2%.
A7 2026-08-20 metr Expenditure Horizon: Measuring Optimization Ability, with an…
https://metr.org/blog/2026-07-21-expenditure-horizon/
major (score, safety, scope)
v137→v322
Added approx. 200 lines: human returns on NanoGPT (approx. $2,500 per 1%), agent runs for six models with expenditure horizons $600-$3,300, maintainer mergeability review, appendices.
A8 2026-08-20 metr Many SWE-bench-Passing PRs Would Not Be Merged into Main
https://metr.org/notes/2026-03-10-many-swe-bench-passing-prs-would-not-be-merged-into-main/
minor
v145→v324
Added Summary paragraph (grader overstates time horizon; trend unsupported) and seven methodological footnotes (Epoch harness lag, 2% corrupted patches, 31 pilot patches, blinding caveats).
A9 2026-08-22 poolside Laguna S 2.1
https://huggingface.co/poolside/Laguna-S-2.1
minor (license)
v254→v345
Added paragraph stating Laguna S 2.1 is released under OpenMDW-1.1, fully permissive, with paid support/optimization/indemnification options; counters changed.
A10 2026-08-22 poolside Laguna XS 2.1
https://huggingface.co/poolside/Laguna-XS-2.1
minor (license)
v255→v346
Added paragraph stating release under OpenMDW-1.1 (fully permissive, commercial use allowed) with paid support and indemnification options. Counters and widget reorder incidental.
A11 2026-08-22 openai GPT-5.6 System Card
https://deploymentsafety.openai.com/gpt-5-6/gpt-5-6.pdf
minor (score, safety, correction)
v7→v348
New Aug 19, 2026 change-log entry and table fix: GPT-5.5 pass@4 on hard-negative protein binding corrected 0.4% to 1.5% (was pass@1). Rest is page-number reflow.
A12 2026-08-22 xai Model Card: Grok 4.6
https://media.x.ai/v1/website/card-4p6-4cd2dc57.pdf
major (score, safety, correction)
v236→v357
Revision 2026-08-17 adds changelog, PartBench, DeepSearchQA, BixBench, KernelBenchInternal v1.1, renumbered sections; corrects HackerBench, self-harm, MASK, LAB scores.
A13 2026-08-24 google deepmind Gemini 3.7 Flash Model Card
https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-3-7-Flash-Model-Card.pdf
minor (safety)
v277→v358
Frontier Safety section changed from ‘FSF Report will be published shortly’ to ‘is available here’.
A14 2026-08-24 openai GPT-5.6 Preview System Card
https://deploymentsafety.openai.com/gpt-5-6-preview/gpt-5-6-preview.pdf
minor (score, safety, correction)
v214→v360
Added dated Change log (Aug 19, 2026) and inline note correcting GPT-5.5 pass@4 on hard-negative protein binding from 0.4% to 1.5%; table updated.
A15 2026-08-25 tencent hunyuan UI-Mate-27B Model Card
https://huggingface.co/tencent/UI-Mate-27B
major (score, scope)
v280→v376
Demonstration-guided mode (description, results table, pipeline) removed and assigned to a separate UI-Mate-democua-27B checkpoint in a new checkpoint table.
A16 2026-08-25 tencent hunyuan EVIE-Preview-4.5B Model Card
https://huggingface.co/tencent/EVIE-Preview-4.5B
major (score)
v281→v377
Card rewritten around ‘Rank #1’ claims: new ViDoRe V3 table with 1,792-token tier (65.36), revised V1+V2 and per-domain scores, Index Cost section.
A17 2026-08-25 nvidia NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 Model Card
https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16
minor
v282→v378
Added one sentence in the training-data section pointing to the Public Summary of Training Content. Download counter and eval-widget rows (Terminal-Bench 2.1 24.58) are furniture.
A18 2026-08-25 inclusion ai Ling-3.0-tiny Model Card
https://huggingface.co/inclusionAI/Ling-3.0-tiny
minor (correction)
v283→v379
Activated parameter count reverted from 1.4B (Non-emb 1.14B) back to 1.3B in the same four places. Total 7.9B unchanged.
A19 2026-08-25 openai GPT-5.5 System Card
https://deploymentsafety.openai.com/gpt-5-5/gpt-5-5.pdf
major (score, safety, correction)
v8→v387
Change log dated August 19, 2026 added; GPT-5.5 pass@4 on hard-negative protein binding corrected from 0.4% to 1.48%; page numbers shifted.
A20 2026-08-26 nvidia NVIDIA Nemotron 3 Nano Omni 30B-A3B Reasoning Model Card
https://huggingface.co/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16
minor
v76→v398
One line added to the dataset section pointing to a Public Summary of Training Content. Remaining changes are download and Spaces counters.
A21 2026-08-26 nvidia NVIDIA Nemotron Parse 2.0 Model Card
https://huggingface.co/nvidia/NVIDIA-Nemotron-Parse-2.0
major (license)
v84→v406
License changed from NVIDIA Open Model License Agreement to OpenMDW-1.1; third-party software notice added; Quick Start and vLLM instructions rewritten.
A22 2026-08-27 openai GPT-Rosalind-5.5 System Card
https://deploymentsafety.openai.com/gpt-rosalind-5-5/gpt-rosalind-5-5.pdf
minor (score, safety, correction)
v31→v419
Change log (Aug 19, 2026) added: GPT-5.5 pass@4 on hard-negative protein binding corrected from 0.4% to 1.48%; table updated; rest repagination.
A23 2026-08-28 google deepmind Gemini Omni Flash Model Card
https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-Omni-Flash-Model-Card.pdf
minor (scope)
v36→v432
Marked Last Updated August 2026; card now also covers Gemini Omni 1.1 Flash; removed paragraph promising T2VA/I2VA/R2VA/editing/image-gen evaluations at API rollout.
A24 2026-09-03 inclusion ai UI-Venus-2-9B
https://huggingface.co/inclusionAI/UI-Venus-2-9B
major (score, license, safety, correction)
v429→v479
Result tables rebuilt with baselines and new benchmarks; safety table now OSHarm+OSBlind; Release Status added, license pending; web claim cut to 4,000+ domains.
A25 2026-09-03 google deepmind Gemini 3.6 Flash Model Card
https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-3-6-Flash-Model-Card.pdf
minor
v14→v480
Intended Usage list adds ‘complex video reasoning’ and replaces ‘multi-week enterprise processes’ with ‘enterprise workflows’. No other section changed.
A26 2026-09-03 google deepmind Gemini 3.5 Flash-Lite Model Card
https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-3-5-Flash-Lite-Model-Card.pdf
minor
v439→v481
Benefit and Intended Usage now lists complex video reasoning among use cases; one sentence split. Nothing else changed.
A27 2026-09-05 inclusion ai Ling-3.0-flash Model Card
https://huggingface.co/inclusionAI/Ling-3.0-flash
minor
v186→v491
vLLM install instructions switched from the inclusionAI vllm-ling-v3 fork to upstream vllm-project/vllm. HF use-instructions block, counters and eval-widget rows are furniture.
A28 2026-09-05 nvidia NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 Model Card
https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16
minor
v270→v494
Training-data section gained one line pointing to a Public Summary of Training Content; rest is HF download/Spaces counters and eval-widget reordering.
A29 2026-09-05 nvidia Cosmos 3: Omnimodal World Models for Physical AI — Cosmos3-S…
https://huggingface.co/nvidia/Cosmos3-Super
minor
v54→v496
One sentence added to the training-data section pointing to a Public Summary of Training Content; download counter changed.
A30 2026-09-05 nvidia NVIDIA-Nemotron-3-Super-120B-A12B-BF16 Model Card
https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16
minor
v53→v495
Spec table gains Speculative Decoding row (MTP head, MTPv2 separate checkpoint); training section links NVIDIA’s Public Summary of Training Content.
A31 2026-09-05 nvidia Cosmos 3: Omnimodal World Models for Physical AI — Cosmos3-E…
https://huggingface.co/nvidia/Cosmos3-Edge
minor
v273→v497
Added ‘Update August 25, 2026’ notice that checkpoint, runtime defaults, examples and benchmarks were updated; link to Public Summary of Training Content.
A32 2026-09-05 openai OpenAI and Hugging Face partner to address security incident…
https://openai.com/index/hugging-face-model-evaluation-security-incident/
minor (safety)
v104→v499
Pending-review sentence replaced by an ‘Update on August 26, 2026’ linking to published findings on the Hugging Face incident.
A33 2026-09-06 palisade research Language Models Can Autonomously Hack and Self-Replicate
https://palisaderesearch.org/blog/self-replication
replaced
v194→v508
Blog URL now serves a meta-refresh stub ‘Redirecting… Click here’ pointing to palisaderesearch.org/research/self-replication; article text no longer captured.
A34 2026-09-06 palisade research Technical Report: Shutdown Resistance in Large Language Mode…
https://palisaderesearch.org/blog/shutdown-resistance-on-robots
replaced (safety)
v195→v509
Page body (shutdown-resistance-on-a-robot report: 3/10 physical, 52/100 simulated trials) replaced by a two-line redirect stub: Redirecting… Click here if you are not redirected.
A35 2026-09-06 google deepmind Gemini 3.7 Flash Model Card
https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-3-7-Flash-Model-Card.pdf
minor
v441→v513
Description adds ‘support for agentic video understanding’; intended use cases add ‘complex video reasoning’.
A36 2026-09-09 nvidia GR00T-H model card
https://huggingface.co/nvidia/GR00T-H
minor (license)
v328→v525
Header notice now labels this the N1.6 non-commercial version and points to GR00T-H-N1.7 as the commercially usable version.
A37 2026-09-10 tencent hunyuan Hy4 preview Model Card
https://huggingface.co/tencent/Hy4-preview
minor
v433→v533
Deployment section dropped build-vLLM-from-source steps and the explicit vllm serve MTP command, pointing to recipes and the prebuilt image instead.
A38 2026-09-12 openai GPT-6 Astra System Card
https://deploymentsafety.openai.com/gpt-6-astra/gpt-6-astra.pdf
major (safety)
v500→v551
Adds Sept 9 change log; renames 8.7 to Verbalized Metagaming and Oversight Gaming, drops metric plot, adds CoT examples; Alignment section adds generalization caveats.
A39 2026-09-14 openai Introducing GPT-Rosalind for life sciences research
https://openai.com/index/introducing-gpt-rosalind/
major (license, safety)
v465→v562
Sept 11, 2026 update: exits research preview, global trusted access, pricing from Oct 5. Removed free-credits paragraph, customer quote, and preview-terms compliance clause.
A40 2026-09-14 openai OpenAI Daybreak: Trusted Access for Cyber — Overview
https://help.openai.com/en/articles/20001258-openai-daybreak-trusted-access-for-cyber-overview
minor (safety)
v466→v563
Two FAQ entries added: reduced refusals on Astra only for Daybreak Red, not Blue; Daybreak Red is org-only and not open to individual applicants.
A41 2026-09-15 deepseek DeepSeek-V4-Flash-Vision-Exp
https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-Vision-Exp
minor
v473→v564
Added a ‘How to Run with vLLM’ section with a 4xGB300 docker serving command using DSpark speculative decoding. Widget and nav changes are furniture.
A42 2026-09-16 anthropic An alignment assessment of recent cybersecurity incidents
https://www.anthropic.com/research/alignment-assessment-cybersecurity-incidents
minor (safety, correction)
v540→v579
Sept 10 correction note added; PyPI removal window changed from approx. 90 minutes to under an hour; research model penetrated one neighbor, not several. Related-content footer changed.
A43 2026-09-16 deepseek DeepSeek-V4.1-Flash
https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash
minor
v541→v580
Agentic evaluation methodology note rewritten: per-benchmark harness assignments, no-network condition for one benchmark, sampling settings stated. No score changed.
A44 2026-09-16 anthropic Real-time cyber safeguards on Claude Opus and Sonnet
https://support.claude.com/en/articles/14604842-real-time-cyber-safeguards-on-claude-opus-and-sonnet
major (safety, scope)
v542→v581
Cyber Verification Program now available via ‘Claude on Google Cloud’ (was unavailable on Vertex); new section requires consenting to the Advanced AI Safety Addendum and setting PublisherModelConfig data-retention fields per model.
A45 2026-09-16 far ai Revisiting Frontier LLMs’ Attempts to Persuade on Extreme To…
https://www.far.ai/blog/revisiting-attempts-to-persuade
major (score, safety, scope)
v193→v583
Two sections added: reasoning-level effect (Gemini 3 harmful persuasion approx. 90% low vs approx. 25% high; GPT-5.1 and Opus 4.5 approx. 0%) and Gemini 3 Flash results.
A46 2026-09-23 anthropic Introducing Claude Fable 5.1 and Claude Mythos 5.1
https://www.anthropic.com/claude-fable-and-mythos-5-1
minor (safety)
v488→v605
Added methodology note: Fable scored zero on OSWorld/AutomationBench where production safeguards intervened; other cyber/bio interventions completed by Opus 4.8/Opus 5. Minor table cell formatting.
A47 2026-09-23 stepfun StepAudio 3 Realtime Technical Report
https://arxiv.org/abs/2609.14005
major
v586→v608
arXiv revision v2 (19 Sep 2026) posted with a submission history added. The abstract and its scores are unchanged; the PDF shrank from 937 KB to 367 KB.
A48 2026-09-23 nvidia Cosmos 3: Omnimodal World Models for Physical AI — Cosmos3-S…
https://huggingface.co/nvidia/Cosmos3-Super
major (scope)
v496→v613
Removed Cosmos3-Nano, Nano-Policy-DROID, Super-Image2Video and Super-Text2Image from Model Versions and parameter lists; card now covers Cosmos3-Super (64B) only.
A49 2026-09-23 nvidia Cosmos 3: Omnimodal World Models for Physical AI — Cosmos3-E…
https://huggingface.co/nvidia/Cosmos3-Edge
major (scope)
v497→v614
Removed eight other Cosmos3 variants (Edge-Policy-DROID, Super 4-step distills, 05/31 Nano/Super release block) from model versions and parameter lists; card now Cosmos3-Edge only.
A50 2026-09-24 openai GPT-6 Astra System Card
https://deploymentsafety.openai.com/gpt-6-astra/gpt-6-astra.pdf
major (score, safety, scope, correction)
v551→v620
Added approx. 35-page Appendix A covering GPT-6 Sol and Luna. Corrected Astra HealthBench scores after a misconfiguration. Added updated alignment-eval results. Three new changelog entries dated Sept 22.
A51 2026-09-25 inclusion ai Ling-3.0-tiny Model Card
https://huggingface.co/inclusionAI/Ling-3.0-tiny
minor
v379→v634
vLLM install now uses upstream vllm-project/vllm instead of inclusionAI fork; added ‘Training content summary’ section pointing to training-content disclosure document.
A52 2026-09-26 openai OpenAI Daybreak: Trusted Access for Cyber — Overview
https://help.openai.com/en/articles/20001258-openai-daybreak-trusted-access-for-cyber-overview
major (safety, scope)
v563→v646
Tier table rebuilt: Blue gives reduced refusals on GPT-5.5/5.6 Sol/GPT-6 Sol/Luna; Red adds GPT-5.5-Cyber and Astra; GPT-5.6-Cyber needs extra approval; FIDO2 key requirement added.
A53 2026-09-28 inclusion ai Ling-2.6-1T model card
https://huggingface.co/inclusionAI/Ling-2.6-1T
minor
v530→v659
Added short Training content summary section linking public training-content disclosure; HF download counter changed.
A54 2026-09-28 inclusion ai Ling-2.6-flash model card
https://huggingface.co/inclusionAI/Ling-2.6-flash
minor
v531→v660
Added ‘Training content summary’ section pointing to a public training-content disclosure document; states it does not replace docs, usage terms or license.
A55 2026-09-28 inclusion ai Ring-2.6-1T model card
https://huggingface.co/inclusionAI/Ring-2.6-1T
minor
v532→v661
Added short Training content summary section linking public training-content disclosure; other changes are HF download count and eval-widget row reorder.
A56 2026-09-28 inclusion ai Ling-3.0-flash-VL
https://huggingface.co/inclusionAI/Ling-3.0-flash-VL
minor (correction)
v528→v662
Context window corrected from 1M to 256K tokens in two places; dropped –mem-fraction-static serving flag; added Training content summary section pointing to disclosure document.
A57 2026-09-29 anthropic Real-time cyber safeguards on Claude Opus and Sonnet
https://support.claude.com/en/articles/14604842-real-time-cyber-safeguards-on-claude-opus-and-sonnet
minor (safety, scope)
v581→v671
Scope note now excludes Claude Opus 5.5 and Sonnet 5.5 from these cyber safeguards; says Cyber Verification Program will expand to Opus/Sonnet 5.5 and Mythos.
A58 2026-09-29 google deepmind Gemini 3.8 Audio (Live, Live Extended Thinking) Model Card
https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-3-8-Audio-Model-Card.pdf
major (safety, scope)
v584→v677
Card expanded to add Gemini 3.8 Flash TTS and Flash-Lite TTS plus Live Avatar video output; added child safety evaluations; frontier safety statement extended; Vertex AI renamed.
A59 2026-09-29 redwood research Astra is much better at reasoning with filler tokens than pr…
https://blog.redwoodresearch.org/p/astra-is-much-better-at-reasoning
minor
v623→v686
Added a 10-shot prompting filler-token result: Astra improves significantly and other models do not. The table is linked to LessWrong, not included.
A60 2026-10-02 inclusion ai Ling-3.0-flash Model Card
https://huggingface.co/inclusionAI/Ling-3.0-flash
minor
v607→v713
Added three-line ‘Training content summary’ section linking a public training-content summary; states it does not replace technical docs, terms, or license.
A61 2026-10-03 openai GPT-6 Astra System Card
https://deploymentsafety.openai.com/gpt-6-astra/gpt-6-astra.pdf
major (score, safety, scope)
v620→v720
New approx. 20-page Appendix B on dots (always-on GPT-6 Astra agents): prompt-injection red-teaming, alignment across time budgets, monitor evasion, preparedness, safeguards; changelog entry Sept 29.
A62 2026-10-03 metr Red-Teaming Anthropic’s Internal Agent Monitoring Systems
https://metr.org/blog/2026-03-25-red-teaming-anthropic-agent-monitoring/
major (score, safety)
v73→v725
Replaced one-line update pointer with new ‘Results from the Exercise’ section quoting monitoring bypasses, monitor jailbreaks, and attack-ideation eval (best 59% vs approx. 90% human).
A63 2026-10-05 openai OpenAI Daybreak: Trusted Access for Cyber — Overview
https://help.openai.com/en/articles/20001258-openai-daybreak-trusted-access-for-cyber-overview
major (safety)
v646→v742
New ‘Set up Daybreak as an individual’ checklist section; hardware-key rule relaxed (other passkeys may coexist with a FIDO2 key); Bedrock alias-unavailability note removed.
A64 2026-10-07 anthropic Introducing the Life Sciences Verification Program
https://www.anthropic.com/news/life-sciences-verification-program
minor (correction)
v590→v752
Links reworded from ‘apply’/‘submit an application’ to ‘join waitlist’/‘register interest’; changelog dated Sept 30, 2026 added; related-content footer rotated.
A65 2026-10-07 mistral Model Card: Leanstral 1.5
https://docs.mistral.ai/models/leanstral-1-5
minor
v244→v753
Status changed from Public Preview to Deprecated with notice ‘being phased out’ and deprecation date 9/29/2026; Free price and duplicated spec block removed.
A66 2026-10-08 anthropic Real-time cyber safeguards on Claude Opus and Sonnet
https://support.claude.com/en/articles/14604842-real-time-cyber-safeguards-on-claude-opus-and-sonnet
major (safety, scope)
v671→v774
Article rewritten: Cyber Verification Program now has three access tiers covering Opus 5.5, Sonnet 5.5, Mythos 5.1; new application, security, eligibility, FAQ sections.
A67 2026-10-08 google deepmind Gemini 3.8 Audio (Live, Live Extended Thinking) Model Card
https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-3-8-Audio-Model-Card.pdf
minor
v677→v779
Added Gemini Enterprise Agent Platform as a distribution channel for 3.8 Flash TTS and Flash-Lite TTS; rest is re-rendered PDF page numbers.
A68 2026-10-08 anthropic Claude Mythos
https://www.anthropic.com/claude/mythos
major (safety)
v678→v780
Cyber Verification Program now covers Opus 5.5, Sonnet 5.5, Mythos 5.1; US-only and ‘near future’ caveats removed; Life Sciences program no longer invite-only beta.
A69 2026-10-11 anthropic Project Glasswing: Securing critical software for the AI era
https://www.anthropic.com/glasswing
major (safety, scope)
v463→v802
Added October 6, 2026 update: expanded Cyber Verification Program with three access tiers covering Opus 5.5, Sonnet 5.5, Mythos 5.1; Glasswing members move to Specialized tier.
A70 2026-10-11 epoch ai Long-context latency scales quadratically for GPT-5.6 but ne…
https://epoch.ai/publications/long-context-latency-scaling-gpt-vs-claude
major (score, scope)
v556→v804
New Appendix H adds GPT-6 Sol and GPT-6.1 Sol TTFT measurements and fitted-coefficient table; appendix headings renamed, Code moved to Appendix I.
A71 2026-10-14 mistral Introducing Mistral Large 4
https://mistral.ai/news/mistral-large-4/
minor (correction)
v769→v829
Stated active parameter count for Mistral Large 4 silently changed from 49 billion to 52 billion.
A72 2026-10-14 anthropic Real-time cyber safeguards on Claude Opus and Sonnet
https://support.claude.com/en/articles/14604842-real-time-cyber-safeguards-on-claude-opus-and-sonnet
minor (license, safety)
v774→v832
Added ‘Building client-facing products’ section: CVP grants cover internal security work only; productizing (e.g. running on customer code) needs separate approval.
A73 2026-10-15 stepfun StepAudio 3 Realtime Technical Report
https://arxiv.org/abs/2609.14005
major
v608→v839
arXiv abstract page now lists a v3 revision (12 Oct 2026, 926 KB); abstract text unchanged, so the paper’s actual changes aren’t visible here.
A74 2026-10-16 redwood research Untrusted advice for AI control: Short, strong advice signif…
https://blog.redwoodresearch.org/p/untrusted-advice-for-ai-control-short
minor (safety, correction)
v512→v849
TL;DR safety claim softened: ‘we argue that this achieves near-maximal safety’ became ‘we speculate that with a sufficiently strict character limit… would achieve’ it.
A75 2026-10-17 anthropic Investigating unintended model actions in our evaluations an…
https://www.anthropic.com/research/investigating-unintended-model-actions
major (safety, correction)
v794→v861
Added Oct 12 correction: URL-length limits on web fetch now described as reducing data-exfiltration risk, not preventing SQL/command injection. Copy edits; related links rotated.
A76 2026-10-18 openai OpenAI Daybreak: Trusted Access for Cyber — Overview
https://help.openai.com/en/articles/20001258-openai-daybreak-trusted-access-for-cyber-overview
major (safety, scope)
v742→v869
Tier list drops the ‘Red with additional model approval’ (GPT-5.6-Cyber) row, though the access matrix keeps its column; Red now ‘cyber models like GPT-5.5-Cyber’; GPT-6.1-Sol, Astra Minor added.
A77 2026-10-21 google deepmind Gemini 4 Argon — Safety & Security Report
https://storage.googleapis.com/deepmind-media/gemini/gemini_4_safety_report.pdf
minor (score, safety, correction)
v834→v901
Footnote qualifies ‘we avoid training against our monitors’, citing early-RL reward-hacking penalty over thoughts (Sec. 3.4.5.2); eval-awareness rates revised; ‘mitigated as if at CCL’ dropped.
A78 2026-10-21 anthropic Claude Mythos
https://www.anthropic.com/claude/mythos
minor (safety)
v780→v904
Safety sentence hedged: models are tested ‘in an effort to ensure’, not ‘to ensure’, they meet Anthropic’s standards; announcements list reformatted.
A79 2026-10-23 metr Early work on monitorability evaluations
https://metr.org/blog/2026-01-19-early-work-on-monitorability-evaluations/
minor
v589→v921
New footnote: human baseline times/estimates are for side tasks solved standalone, not within side-task setup; later footnotes renumbered 4-8 to 5-9.
A80 2026-10-25 epoch ai Claude overperforms at software engineering and underperform…
https://epoch.ai/data-insights/claude-ds-eci
minor
v152→v958
Added a Limitations section: domain-specific ECI is relative-only, and models have uneven math vs SWE benchmark counts (avg 3.4 vs 3.2, max 4:2).
A81 2026-10-25 epoch ai Disclosed CVEs: July Reached 5x the Pre-Mythos Record
https://epoch.ai/data-insights/cve-severity-spike-july-2026
major (score, safety, correction)
v154→v960
Data section filled (21 source organizations listed); Assumptions and limitations added; pre-Mythos monthly CVE record revised upward from approx. 440 to approx. 490.
A82 2026-10-25 anthropic Training a Misaligned Reward Seeker
https://alignment.anthropic.com/2026/reward-seeker/
minor (safety)
v474→v962
One related-work paragraph added citing Hojmark et al. (2026) on measuring reward seeking via grader-belief SDF in OpenAI o3 and gpt-oss-120b.
A83 2026-10-26 epoch ai GPT-6 Astra leads on math benchmarks, but not on software en…
https://epoch.ai/data-insights/astra-eci-breakdown
major
v592→v970
Data section filled in (cutoff date, benchmark lists, per-model benchmark-count table) and three limitations bullets added on small benchmark counts, refit drift, and relative ECIs.
A84 2026-10-26 meta Muse Wonder Model Card
https://research.meta.ai/static/muse-wonder-model-card
minor (score, safety, correction)
v815→v973
ExploitGym result for Muse Wonder Full corrected from 92 (21 intended) to 20 intended flags after harness fix; dated update note (23 Oct 2026) replaces footnote.
A85 2026-10-27 inclusion ai LLaDA2.2-flash model card
https://huggingface.co/inclusionAI/LLaDA2.2-flash
minor
v763→v980
Deployment/SGLang placeholder (‘coming soon’) replaced with full SGLang install and launch_server instructions, including JointThresholdInDel decoding defaults; download counter bumped.
A86 2026-10-27 anthropic Real-time cyber safeguards on Claude Opus and Sonnet
https://support.claude.com/en/articles/14604842-real-time-cyber-safeguards-on-claude-opus-and-sonnet
major (license, safety)
v889→v990
CVP productization rules added: managed-service applications open, products interest-only; approval now needs workspace activation; Google Cloud requires Advanced AI Safety Addendum and data sharing.
A87 2026-10-28 google deepmind Fairwind Program — Google DeepMind
https://deepmind.google/fairwind-program/
minor (safety)
v835→v1001
Access-request route changed from ‘fill in our form’ to ‘contact your Cloud account team’, adding a Google Cloud get-started link; one subheading dropped by extractor drift.
A88 2026-10-29 nvidia NVIDIA Nemotron 3.5 Bolt 74B-A9B 24GB — Model Card
https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB
minor (score)
v858→v1017
Benchmark table: Gemma4 comparison column relabeled 26B-A4B-IT-BF16; missing Gemma4 MCP-Atlas score filled in as 48.4. Download counter also changed.
A89 2026-10-30 metr Red-Teaming OpenAI’s Internal Agent Monitoring Systems
https://metr.org/blog/2026-10-22-red-teaming-openai-agent-monitoring/
minor (safety, correction)
v939→v1022
Remediation claim softened from ‘all production-safeguard vulnerabilities patched’ to ‘most patched, one fix in progress’; publication-delay footnote and Oct 29 update log added.
A90 2026-10-31 openai OpenAI Daybreak: Trusted Access for Cyber — Overview
https://help.openai.com/en/articles/20001258-openai-daybreak-trusted-access-for-cyber-overview
major (safety, scope)
v869→v1027
New ‘Using GPT-6.1 Terra with Daybreak’ section; Terra requires explicit Daybreak Red enablement, no default to highest tier; Codex notes updated; typo/link-label fixes.
A91 2026-11-01 securebio SecureBio’s GPT-6 Astra Pre-Release Testing Report
https://securebio.substack.com/p/securebio-gpt-6-astra-pre-release
major (score, safety, correction)
v569→v1030
Correction notice added: GPT-6 Astra WCB score revised from 61.6% to 59.2% (web search had been enabled); gap to GPT-5.6 Sol now approx. 8 points.
A92 2026-11-02 mistral Introducing Mistral Large 4
https://mistral.ai/news/mistral-large-4/
minor
v829→v1040
API pricing cut: input $1.36 to $0.7 per M tokens, output $4.18 to $2.1 per M tokens; no other text changed.
A93 2026-11-03 anthropic Real-time cyber safeguards on Claude Opus and Sonnet
https://support.claude.com/en/articles/14604842-real-time-cyber-safeguards-on-claude-opus-and-sonnet
minor
v990→v1045
Removed past Oct 14 CVP webinar notice; added step-by-step Console workspace-creation instructions for provisioning approved cyber-program access.
A94 2026-11-03 mistral Mistral Medium 3.5 model card
https://docs.mistral.ai/models/model-cards/mistral-medium-3-5-26-04
major
v837→v1046
Status badge changed GA to Deprecated; pricing ($1.5/$7.5 per M tokens) removed; deprecation notice added with date 10/31/2026 and replacement Mistral Large 4.
A95 2026-11-03 anthropic Claude Mythos
https://www.anthropic.com/claude/mythos
major (safety)
v904→v1049
New page header and FAQ section added explaining Mythos 5.1 misuse risks, Fable 5.1 as the safeguarded variant, and gated access via verification programs.
A96 2026-11-03 xiaomi MiMo-V2.5 Model Card
https://huggingface.co/XiaomiMiMo/MiMo-V2.5
major (license, safety)
v1003→v1051
New Use Restrictions section (unlawful, military, malicious-cyber, unauthorized personal-data, unsupervised high-risk autonomous use) plus legal contact; states it does not amend the licence. Widget rows reshuffled.
A97 2026-11-03 xiaomi MiMo-V2.5-Pro Model Card
https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro
major (license, safety)
v905→v1050
New Use Restrictions section (no unlawful, military, malicious-cyber, unauthorized personal-data, or unsupervised high-risk autonomous use) plus legal contact; HF widget rows reshuffled.
A98 2026-11-04 fangcun ai Evaluating DeepSeek V4.1 Flash-1017 for loss of control and …
https://fangcunleap.com/public/reports/fangcun_eval_deepseek_v4_1_flash_1017.pdf
minor (safety, correction)
v1011→v1056
PINN Cholesky-bug explanation in PaperBench case study rewritten (mechanism corrected, conclusion same); Table A.1 now lists V4 Flash accessed via INFINI-AI GenStudio, not DeepSeek API.
A99 2026-11-04 openai GPT-6 Astra System Card
https://deploymentsafety.openai.com/gpt-6-astra/gpt-6-astra.pdf
minor (score, safety, correction)
v720→v1057
Added Oct 30, 2026 change-log entry: SecureBio corrected GPT-6 Astra WCB score from 61.6% to 59.2% (web search had been enabled); Preparedness text updated.
A100 2026-11-06 mistral Le Chonk: Built for Coders, Trusted by Defenders (Mistral La…
https://mistral.ai/news/mistral-large-4-oss/
minor
v1025→v1068
Launch text updated from ‘technical report coming soon’ to report now available; trailing product tile with $0.7/$2.1 per M token pricing removed.
A101 2026-11-10 anthropic Claude Mythos
https://www.anthropic.com/claude/mythos
major (score, license, safety, scope)
v1049→v1116
Page retitled Mythos 5.1 to 5.2; pricing cut to $8/$40 plus $16/$80 fast mode; safeguards section shortened; data-retention/ZDR terms changed; 15-row benchmark table added.
A102 2026-11-11 nvidia NVIDIA Isaac GR00T N1.7 Model Card
https://huggingface.co/nvidia/GR00T-N1.7-3B
minor
v108→v1133
One sentence added under License: ‘Built on NVIDIA Cosmos’, naming nvidia/Cosmos-Reason2-2B as the vision-language backbone. Download counter changed.
A103 2026-11-11 nvidia Cosmos-H-Dreams Model Card
https://huggingface.co/nvidia/Cosmos-H-Dreams
minor
v722→v1134
Added body note pointing to nvidia/Cosmos3-H-Dreams as the latest version; HF ‘Instructions to use’ block and download counter also changed (furniture).
A104 2026-11-11 xiaomi MiMo-V2.6-Pro-RL model card
https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL
major (license, safety)
v792→v1135
New ‘Use Restrictions’ section added: bans unlawful, military, malicious-cyber, unauthorized data, and unsupervised high-risk autonomous use; legal contact added. Plus HF widget churn.
A105 2026-11-11 nvidia NVIDIA Nemotron 3.5 Super VL 120B-A12B — Model Card
https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16
minor
v856→v1138
Serving section rewritten: SGLang added to hardware table, MTP draft tokens 3->5 on H100, new SGLang recipes for 4xH100 and 2xGB200 with nemotron3-5-super image; plus HF widget furniture.
A106 2026-11-11 openai Bringing GPT-Rosalind to academic research labs (Rosalind Ac…
https://openai.com/index/introducing-the-rosalind-academic-plan/
minor
v859→v1140
Added a four-step application walkthrough (explore plan, sign in and apply, complete verification and sign agreement, start research); hyphen in ‘GPT-Rosalind’ changed to non-breaking.
A107 2026-11-11 google deepmind Gemini 4 Carbon: our next frontier model
https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/
minor (safety)
v1018→v1141
Safety section clarified: ‘our Safety and Security report’ now reads ‘our Safety and Security report for Argon’; no other text changed.
A108 2026-11-11 google deepmind Gemini 4 Carbon — Model evaluation: Approach, methodology & …
https://storage.googleapis.com/deepmind-media/gemini/gemini_4_carbon_model_evaluation.pdf
minor (safety)
v1019→v1142
Methodology notes reworded: FrontierSWE renamed FrontierSWE 2, CWE-bench attribution adds Artificial Analysis, ExploitGym ‘mitigations off’ clarified as benchmark mitigations, results header adds Fairwind Program caveat.
A109 2026-11-13 thinking machines Inkling 2 Small Model Card
https://thinkingmachines.ai/model-card/inkling-2-small/
major (score, license, safety)
v1073→v1163
Open-weights release: Apache 2.0 license added, HF download/hardware section, approx. 20 new benchmark rows incl. HoneyBench safety row, many scores filled or revised.
A110 2026-11-13 metr Assessment of AI R&D acceleration at Anthropic
https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/
minor (safety, correction)
v1075→v1165
Added Nov 6 Update log; AEF-1 checklist item 1.1 (sufficient technical access) corrected Yes to No; template notes removed; footnote arrow glyphs changed.