Skip to content

Search stories, models, companies…Subscribe

Topic · Concept

System cards

System cards are the documents AI labs publish alongside major models to describe capabilities, safety testing and deployment decisions. They are among the most detailed public disclosures labs make, but they are written by the developers themselves and vary widely in what they include.

10 stories · updated Oct 18

Follow via RSS Get System cards alerts

Key facts

Checked Oct 8

Model cards (origin): Mitchell et al., "Model Cards for Model Reporting," Oct 5, 2018
Anthropic definition: "Document the capabilities, safety evaluations, and responsible deployment decisions for Claude models"
Latest Anthropic system cards: Claude Haiku 5.5 (October 7, 2026), Claude Sonnet 5.5 (September 28, 2026) and Claude Opus 5.5 (September 22, 2026)
Latest OpenAI safety publication listed: "GPT-6 Sol and GPT-6 Luna: October 2026 update," Oct 7, 2026

Start here

Research · ExplainerHow to read a system card: a builder's guide to AI safety documentationSystem cards run to hundreds of pages. What the sections mean, which safety frameworks they report against, how to read eval tables critically, and a checklist.Oct 9·7 min read

Research · Analysis"AI models can provably hide arbitrary capabilities": what the paper proves, and what it means for AI safety evaluationsAn October 2026 paper by researchers from Oxford and others shows how to lock a hidden skill inside a model's weights with cryptography, so that auditors with full weight access can't trigger or identify it. What 'provably' means here, the assumptions behind it, and why it matters for system cards.Oct 16·7 min read

Latest on System cards

Research · AnalysisGLM-5.3 update-aware reasoning: what Apollo Research found when Z.ai's model reasoned about its own trainingApollo Research reported on October 15, 2026 that Z.ai's GLM-5.3 sometimes reasons, unprompted, about how its actions could become training data and shape future versions of itself. Here's what the GLM-5.3 update-aware reasoning finding shows, how rare it was, how Apollo measured it, and what it doesn't prove.Oct 18·6 min read

Research · Analysis"AI models can provably hide arbitrary capabilities": what the paper proves, and what it means for AI safety evaluationsAn October 2026 paper by researchers from Oxford and others shows how to lock a hidden skill inside a model's weights with cryptography, so that auditors with full weight access can't trigger or identify it. What 'provably' means here, the assumptions behind it, and why it matters for system cards.Oct 16·7 min read

Policy · AnalysisAnthropic and OpenAI sought control over what independent AI evaluators publish, Reuters reports. Why evaluator independence mattersReuters reported on October 13, 2026, that as recently as October 5, Anthropic and OpenAI were seeking strict editorial control over what outside safety evaluators could publish, weeks after both CEOs pledged evaluators would publish without it. What was reported, what the companies say, and why it matters for every safety claim you read.Oct 14·7 min read

Models · ExplainerOpenAI loop transformers: what's reported about GPT-6.1 Sol and OpenAI's next model, and what a looped transformer isThe Information reported on October 13, 2026, that OpenAI is working through safety issues tied to loop transformers before shipping its next flagship model. What's actually confirmed, what OpenAI's GPT-6.1 Sol safety addendum shows, how the cancelled GPT-6.1 Astra fits in, and how looped transformers work.Oct 14·7 min read

Research · AnalysisWhat the Gemini 4 Argon model card says: safety tests, cyber mitigations, Fairwind access and the limitsGoogle's Gemini 4 Argon model card says the company can't rule out that Argon reached its critical capability levels for biological and cyber risk, so it shipped with safeguards and a safety case. Here's what the card and Google's safety post report, how they compare with the Claude Opus 5.5 system card, and what's missing.Oct 14·9 min read

Models · NewsMeta Muse Wonder and Muse Wonder Full: price, availability and what Artificial Analysis foundMeta released Muse Wonder and Muse Wonder Full on October 12, 2026. Muse Wonder costs $4/$20 per million tokens on Meta's API; Wonder Full is only in paid Muse Code plans. Artificial Analysis scores them 52 and 56 on its Intelligence Index.Oct 14·5 min read

Models · NewsClaude Haiku 5.5: price, 1M context and how it compares with Haiku 4.5Anthropic released Claude Haiku 5.5 on October 7, 2026. It costs $0.10/$0.50 per million tokens for prompts up to 100K, has a 1M-token context window and 128K max output, and is now Claude Code's default Haiku on the Anthropic API.Oct 13·5 min read

Models · NewsGPT-6 reaches every ChatGPT user: what OpenAI shipped, what it costs, and where each model runsFour GPT-6 models in five weeks, split across ChatGPT's Chat and Work tabs, Codex and the API. Here's the verified map, with prices.Oct 9·6 min read

Models · AnalysisWhat the Claude Opus 5.5 system card actually saysAnthropic's system card for Opus 5.5 runs past 220 pages. What it says on bio and cyber risk, sandbox escapes, prompt injection and alignment, and what it leaves out.Oct 9·7 min read

Research · ExplainerHow to read a system card: a builder's guide to AI safety documentationSystem cards run to hundreds of pages. What the sections mean, which safety frameworks they report against, how to read eval tables critically, and a checklist.Oct 9·7 min read

Nothing in this format on this page — try another filter or the next page.