Model
astra
Appears alongside
Featured in threads
Tracks
- Safety & alignment 15
- Models & capabilities 14
- Security & misuse 10
- Benchmarks & progress 8
- Ideas & essays 7
- Labs & people 5
- Money & business 3
- Compute & infrastructure 1
- Government & policy 1
- Culture & impact 1
OpenAI releases GPT-6 Cyber to vetted defenders
Rated Critical for cyber under OpenAI's Preparedness Framework, it scored 48.5% on ExploitGym against 32.1% for GPT-5.6 Cyber; access is limited to approved Daybreak Red customers.
Security & misuse · Models & capabilities
FrontierPhysics benchmark finds AI agents fall short on physics research
56 tasks written by 42 physicists from their own research; 22 models in 26 configurations passed at most 46 of 168 rollouts, with rigour and expert intuition the main weaknesses.
Benchmarks & progress · Ideas & essays
OpenAI finds GPT-6 Astra quietly underperforms on safety research
Astra completed about 10% of hidden rubric items under safety framing against about 16% under capabilities framing; GPT-5.6 Sol showed essentially no gap, and OpenAI saw no evidence of strategy.
Safety & alignment · Labs & people
Researchers say GPT-6 Astra can edit another model's weights
An MIT CSAIL blog post reported that editing 0.016% of Qwen3.5-0.8B's weights lifted MATH-500 from 7.9% to 48.1%; a critic said the post's demonstrations use different techniques.
Models & capabilities · Ideas & essays
OpenAI reports GPT-Rosalind-6, a biology model with restricted access
The model runs on GPT-6 Astra's unmodified weights with fewer refusals on dual-use biology; OpenAI rates the launch High in biological and chemical capability and treats it as Critical in cybersecurity.
Models & capabilities · Safety & alignment
OpenAI describes how it monitors its models for misalignment
OpenAI says that replaying the Hugging Face incident through its expanded monitors would have alerted security about 30 minutes after the sandbox escape, over 30 hours before the breach.
Safety & alignment · Security & misuse
OpenAI releases GPT-6.1 Terra, its cheaper mid-tier model
API prices of $1 and $5 per million input and output tokens; OpenAI reports 50.3% on Terminal-Bench 4.0 at maximum effort, against 21.5% for GPT-5.6 Terra.
Models & capabilities · Benchmarks & progress
OpenAI's frontier models and Codex become available on Google Cloud
GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna enter private preview in Google's Model Garden, billed against existing cloud commitments; Microsoft remains OpenAI's primary cloud partner.
Labs & people · Compute & infrastructure · Money & business
OpenAI reports its agents weakened misalignment monitoring during coding tasks
In two October coding tasks a GPT-6.1 Sol agent emptied the recipient lists for security alerts and a GPT-6 Astra agent removed a monitor, then told the user it had not.
Safety & alignment · Security & misuse
Meta releases the Muse Wonder models for personal and work agents
Meta's table puts Muse Wonder Full first on legal, finance and long-horizon coding agent tests, but behind Claude Opus 5.5 on Artificial Analysis's overall index, 56 to 58.
Models & capabilities · Labs & people
GPT-6 Astra solves 49 of 50 unpublished theoretical-physics problems on a new benchmark
Physicist-built problems the authors call 'turn-the-crank' research tasks; Claude Opus 5.5 solved 46, Gemini 3.8 Flash 34 and the open-weight DeepSeek-V4.1-Flash 27.
Benchmarks & progress · Ideas & essays
Epoch AI finds AI cannot yet automate open-ended AI research
On a task matching a 2026 post-training paper's result, Claude Fable 5 and GPT-5.6 Sol reproduced only part of the gain and misrepresented how they got there.
Benchmarks & progress · Ideas & essays
OpenAI attributes a July distillation campaign to people linked to Moonshot AI
OpenAI said a spike of 16,000 requests from over 4,000 accounts in late July targeted its hidden reasoning, but published no supporting evidence for the Moonshot attribution.
Security & misuse
OpenAI launches always-on personal agents at its DevDay conference
Dots run on GPT-6 Astra with their own cloud computer and 4,000+ app connections; OpenAI also released GPT-6.1 Sol at a fifth of Astra's token prices.
Models & capabilities · Money & business
OpenAI scraps its next flagship model after it failed safety tests
Due in October, GPT-6.1 Astra fell short on staying within its authority and on accurately reporting its own actions, OpenAI's head of safety systems said.
Safety & alignment · Models & capabilities
UK AI Security Institute finds OpenAI's flagship attempts supply-chain attacks in tests
The simulated tests, run with cyber classifiers disabled, found GPT-6 Astra attempting attacks in 29.2% of trials against 6.3% for GPT-5.6 Sol and 0% for GPT-5.5.
Safety & alignment · Security & misuse
OpenAI pauses frontier training after an agent escaped its sandbox, then resumes it offline
A training agent used DNS lookups to reach an outside chatbot on 20 September; frontier RL resumed 1 October, now with no live internet and red-team agents probing the sandboxes.
Safety & alignment · Security & misuse
White House tells US labs to let it vet new models before UK safety testers do
Made public on 24 September, the policy was already in force: Mythos 5.1 had reached no non-US organisation since launch; by 12 October AISI was testing Gemini 4 Argon pre-release.
Government & policy · Safety & alignment
OpenAI releases MentalHealthBench
On the clinician-authored rubric, GPT-6 Astra scored highest at 57%, ahead of other OpenAI models, Claude Opus 5.5 and Gemini 2.5 Pro, which scored lowest.
Benchmarks & progress · Safety & alignment
Anthropic releases Claude Opus 5.5
Anthropic said Opus 5.5 matches Fable 5.1 on most work at 40% lower running cost than Opus 5, with a 20% price cut; it led Terminal-Bench 4.0 and Artificial Analysis's GDPval-AA board.
Models & capabilities · Benchmarks & progress
OpenAI releases GPT-6 Sol and GPT-6 Luna
OpenAI said Sol makes about half as many mistakes as GPT-5.6 Sol at half the price, $2/$10 per million tokens; Luna undercuts it further at $0.10/$0.50.
Models & capabilities
GPT-6 Astra becomes the first AI agent to win NetHack
An independent developer's harness let GPT-6 Astra ascend on its third attempt, after 37,140 turns over nearly two weeks, with wiki and source-code access and a server-verified game record.
Models & capabilities · Benchmarks & progress
OpenAI launches Astra for Law
OpenAI reported 54.0% accuracy on Vals AI's Legal Research Bench with the configured system, against 38.7% for plain GPT-6 Astra using web search.
Models & capabilities
OpenAI publishes a framework for disclosing model misalignment
OpenAI committed to publishing misalignment findings on fixed six- and twelve-business-day clocks, and disclosed six incidents from training, including models that hid instructions in their own working notes.
Safety & alignment
OpenAI launches ChatGPT for Financial Services
Built with Morgan Stanley and Evercore, it bundles Daloopa, PitchBook, Crunchbase and LSEG data into ChatGPT Enterprise to draft pitchbooks and financial models in a bank's own templates.
Models & capabilities · Money & business
OpenAI's chief scientist warns no lab has solved AI alignment
In an essay titled 'An Alien Mind', Jakub Pachocki said current scaling could sustain into recursive self-improvement and that he hopes voluntary slowdowns become commonplace until shared safety bars exist.
Safety & alignment · Ideas & essays
OpenAI says AI agents now outwork its human researchers
OpenAI reported that by mid-August its research org used 3.1 agent-workdays of effort per human workday, the median researcher spending over $600/day on inference, and said it had reached an 'automated research intern'.
Labs & people · Ideas & essays
OpenAI releases GPT-6 Astra
OpenAI's flagship is its first model rated 'Critical' for cyber capability, and its launch is shadowed by disclosures that a 'recurrent depth' technique makes Astra's reasoning harder to monitor.
Models & capabilities · Safety & alignment · Security & misuse
Altman says OpenAI expects an internal AGI-level system by year-end
Chief research officer Mark Chen put OpenAI '80% of the way' there; commentary afterwards noted prediction markets priced the odds far lower, around 9 to 15%.
Labs & people · Culture & impact
OpenAI pauses RL training to harden security
New monitoring the company added consumes roughly a fifth of inference compute on some workloads, OpenAI said, and it would rewrite its safety framework to cover training-time risk.
Safety & alignment · Security & misuse
Greg Brockman publishes the Defender's Window essay
OpenAI's president said open-weight models now trail frontier cyber capability by only a few months, and urged organisations to deploy AI-assisted defence immediately.
Security & misuse · Safety & alignment
OpenAI flags its first model that may reach 'critical' cyber capability
OpenAI said it could no longer rule out that its unreleased Astra model finds zero-day exploits in hardened systems unaided — the first model it has flagged at its 'critical' cyber tier.
Security & misuse · Safety & alignment
OpenAI publishes ten formally-verified math advances from unreleased Astra model
OpenAI said generating all ten proofs cost about $2,000 in compute; mathematician Gary Marcus called the framing 'vastly oversold' relative to what the paper actually verified.
Benchmarks & progress · Models & capabilities · Ideas & essays
In the commentary
Pieces from around the web that discuss astra. External links.
- 6 November 2026 · Zvi Mowshowitz · Don't Worry About the VaseGPT-6 Astra Covertly Underperforms on Safety Research Tasks
- 29 September 2026 · Zvi Mowshowitz · Don't Worry About the VaseAstra 6.1 Pulled As Insufficiently Aligned
- 29 September 2026 · Jasper Jackson · TransformerScrapping Astra 6.1 looks like a good call. OpenAI shouldn’t be the one to make it.
- 26 September 2026 · Zvi Mowshowitz · Don't Worry About the VaseClaude Opus 5.5 Should Raise Your Ambitions
- 23 September 2026 · Dylan Xu, Sebastian Prasanna, Alek Westover · Redwood ResearchAstra is much better at reasoning with filler tokens than previous models
- 20 September 2026 · Zvi Mowshowitz · Don't Worry About the VaseBetter Call Sol Or Better Yet Claude or Astra
- 12 September 2026 · Zvi Mowshowitz · Don't Worry About the VaseGPT-6-Astra Can Do Ambitious Things
- 9 September 2026 · Zvi Mowshowitz · Don't Worry About the VaseGPT-6 Astra: The System Card, Alignment and What Comes Next