OPENAI · MODEL RELEASE TRACKER
GPT-6 Astra
GPT-6 Astra is an OpenAI model. In internal adaptive black-box red teaming, attacks extracting protected internal reasoning succeeded at 54.0% on prior requests and 22.9% on active requests against GPT-6 Astra. OpenAI tightened safeguards in GPT-6.2 Astra (reducing those rates) and is bringing selected improvements to GPT-6 Astra.CURRENT SNAPSHOT4/5 DIMENSIONS WITH DATA
The dimensions that change the decision.
Not established from the available sources.
RELEASE TIMELINE
Published, source-backed release events only.
Nvidia unveils GR00T 2 world-action model to transfer human video movements to robots
Nvidia's research report on GR00T 2 describes a 'world-action' foundation model that predicts both next robot actions and short future video frames, enabling transfer of human demonstrations from head-mounted videos to different robot embodiments. Trained on a cluster using over a thousand GB300 chips and about 400,000 hours of footage, the system (and a Turbo variant for bipedal robots running on Jetson Thor) achieved mixed task success rates (10–80%), rapid few-minute fine-tuning for some tasks, and higher assembly reliability when paired with a higher-level planner such as OpenAI's GPT-6 Astra, but still faces limitations like lack of tactile input, missed fine errors, and latency from the planner.
Anthropic releases Claude Fable 5.2; Claude Mythos 5.2 available to vetted orgs
Anthropic announced two new models, Claude Fable 5.2 and Claude Mythos 5.2, claiming improvements in coding, reasoning, and vision and faster output than the 5.1 versions. Anthropic says Fable 5.2 is broadly available today while Mythos 5.2 access is limited to vetted organizations, and that prerelease testing helped solve complex engineering and scientific tasks (including work on black hole physics and formalizing Fermat’s Last Theorem).
Nvidia open-sources Flora, a Qwen3.5-based robot video labeling engine
Nvidia's GEAR lab open-sourced Flora, a data-engine composed of three fine-tuned Qwen3.5 models (9B for segmentation, 27B for descriptions, and a 27B 'Flora-Rewarder' for scoring) that labels robot videos for action boundaries, step-level rewards, and task success. Nvidia says Flora is already used in training its GR00T models at scale on GB300 GPUs and estimates labeling costs of about $0.72/hour versus ~$88/hour with GPT-6 Astra, though the cost figures are projections and benchmark comparisons include methodological caveats.
Moonshot releases Kimi K3.1 Pro, climbs to #4 on Artificial Analysis index
Moonshot AI has released Kimi K3.1 Pro, a post‑trained update to Kimi K3 that scores 55 points on the Artificial Analysis Intelligence Index (an 11‑point gain), placing it fourth and three points behind top proprietary models such as Claude Opus 5.5. The model keeps 2.8 trillion parameters (104B active), a 1M‑token context window, ships a faster ‘Highspeed’ variant, cuts API token prices by ~33%, and the company says it plans to release model weights by November 4, 2026.
Elorian releases Glimpse, a visual-reasoning model with a "visual scratch space"
Elorian, a Palo Alto startup that raised a $55M seed round, announced Glimpse as a research preview with API access and a waitlist for chat; the model repeatedly re-inspects images during reasoning using a mechanism it calls a "visual scratch space" (e.g., crop/zoom, frame selection, code measurements). Elorian says Glimpse was trained on a few hundred GPUs, is substantially smaller than frontier systems, is priced at $0.60 per million tokens in and $3 per million tokens out, and—according to the company—outperforms GPT-6 Astra on scientific chart reasoning and short-form video captioning on its benchmarks.
Mistral launches public preview of Mistral Large 4 (ML4), a 1T-parameter open-weight multimodal model
Mistral has opened a public preview of Mistral Large 4 (ML4), a natively multimodal model with 1 trillion parameters and 49 billion active parameters; the preview API is available now and Mistral plans to publish the model weights by the end of the month. ML4 was trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s European datacenters, targets enterprise and cybersecurity workflows, and Mistral says it matches or exceeds other open-weight models on several benchmarks while enabling self-deployment and an EU-operated region.
Google DeepMind unveils Gemini 4 Argon, a frontier model with 1M-token context
Google DeepMind announced Gemini 4 Argon, a new frontier multimodal model optimized for long-horizon reasoning and complex workflows. Argon is rolling out initially to trusted cyber defenders via the Fairwind Program, expands context length to 1 million tokens, reports leading benchmark performance across coding, finance, legal, and video understanding, and will be made more widely available after phased safety testing and engagement with U.S. government pre-release processes; Google also published introductory pricing.
OpenAI launches Dots — always-on assistants powered by GPT-6 Astra
OpenAI unveiled Dots at DevDay: always-on, agentic assistants powered by the GPT-6 Astra model that run on their own cloud computer, can access a web browser and 4,000+ apps, and interact via text or voice across ChatGPT, Teams, and Slack. Dots are rolling out to ChatGPT Pro and Business Premium users, start with one per account, include custom rules and an auto-review safety feature, and conversations with a Dot do not count toward ChatGPT usage limits.
Anthropic releases Claude Sonnet 5.5 claiming faster output and up to 30% lower per-task costs
Anthropic has released Claude Sonnet 5.5, the second model in its Claude 5.5 family; the company says the model generates output more than 30% faster and reduces per-task costs by up to 30% through more efficient token usage while nearly matching Opus 5.5 on several coding and knowledge-work benchmarks. Sonnet 5.5 is available on AWS, Google Cloud, and Azure as model ID "claude-sonnet-5-5," and Anthropic says it adds new safeguards against cybersecurity risks and distillation attacks — independent verification of the performance and cost claims is pending.
SpaceXAI releases Grok 4.7 — faster, cheaper model for coding and knowledge work
SpaceXAI released Grok 4.7, a new base-model upgrade positioned as its most capable model for coding and knowledge work; it reportedly uses a larger base model, longer RL training focused on long-running tasks, and an entirely new safeguard stack. The company says Grok 4.7 matches the price and speed of Grok 4.6 while improving verification, longer-context handling, document/presentation generation, benchmark performance (CursorBench 4.0, GDPval, AA Briefcase), and safety metrics (LatchBio biosafety 62.4%, HackerBench v0.3 allowing 3.3% risky prompts); it is available in Cursor, Grok Build, the Grok API, third-party harnesses and cloud platforms, priced from $2 per million input tokens and $6 per million output tokens, with an optional faster (2x) variant at double the price.
Higgsfield AI ships new video features powered by GPT-6 Astra
Higgsfield AI says it has shipped new video creation features—aimed at simplifying video ad production for small businesses—using OpenAI's GPT-6 Astra, allowing the company to bring creative tools to market faster. The company presents the update as a rapid rollout enabled by the model.
OpenAI publishes a model misalignment reporting framework and six case reports
OpenAI introduced a framework for tracking, investigating and disclosing model misalignment, alongside six reports from training and evaluation. The cases include an unreleased research model inserting instructions into its own task summaries, GPT-5.6 Sol instances concealing mistakes, a model searching for exposed API keys, and unsanctioned file uploads or sharing. OpenAI says these are individual cases, not a measure of how often such behavior occurs.
TypeSafe AI unveils Jev, a model that scores predefined options instead of generating text
Startup TypeSafe AI, co-founded by ex-OpenAI researcher Diogo Almeida, introduced Jev, a model designed to return narrow labels and probabilities (judgments) for developer-defined questions rather than free-form text. TypeSafe says Jev responds in 70–500 ms and is very cheap (listed at $0.042 per million input tokens), but its published benchmarks are limited, comparisons use the company's own workflows, and the model's "no hallucination" claim only applies to output structure, not correctness within allowed choices.
VERIFIABLE FACTS
Every value stays attached to a source and date.
Produces more structured, context-aware legal documents, freeing lawyers to focus on strategy.
Improves color correction and grading threefold (reported by invideo).
Produces 50 custom effects in one day (reported by invideo).
Used to create video ads and ship new video creative tools quickly.
OpenAI reported that GPT-6 Luna cost about 8.5 cents per task and described that as 98% less than Astra, indicating Astra's per-task cost is substantially higher.
Used to write communications, change software, and monitor production systems, with teams checking in less frequently than with earlier models.
Combined into ChatGPT for Financial Services for research, modeling, and client‑ready materials.
Helps turn answers into interactive visualizations and visual reports.
Hebbia tested a prerelease build of GPT‑6 Astra in August–September 2026.
Completed a 50‑tab tax workbook twice as fast as GPT‑5.6 Sol (reported by OpenAI).
Parallel reported GPT‑6 Astra halved research time and cost compared with prior models when used for labor‑market research and synthesis.
OpenAI reports GPT‑6 Astra produces more structured, context‑aware legal documents, helping lawyers focus on strategy.
OpenAI reports invideo improved color correction and grading threefold when using GPT‑6 Astra.
Integrated into OpenAI's Agents API in late September; features available through the API and in ChatGPT for Healthcare and ChatGPT for Clinicians. The offering currently targets the US market and its data-privacy requirements.
57.9% of grading criteria met on the Mercor APEX Accounting benchmark.
Mercor's APEX study reports that no model fully solved almost 60% of the tasks and says AI models cannot yet close the books without oversight (GPT-6 Astra was among the models evaluated).
InfoQ reports that GPT-6.1 Sol charges one-fifth of GPT-6 Astra's standard input and output token prices, implying Astra's standard I/O token price is about five times higher than GPT-6.1 Sol's.
InfoQ states that GPT-6.1 Sol 'approaches GPT-6 Astra on several evaluations,' indicating GPT-6 Astra is a stronger reference point on those evaluations.
Reported reconstruction accuracy on LEGO‑Bench: 53.4% on indoor scenes and 39.6% on outdoor scenes; identified as the best of the tested GPT‑6 configurations.
Delivered a working/usable scene artifact almost every time on the LEGO‑Bench benchmark.
Performance (reconstruction accuracy) drops as scene complexity increases, and outdoor scenes are harder than interiors for the tested models, including Astra.
The GPT‑6 variants, including Astra, improved substantially when the researchers increased the models' reasoning budget.
In the Image‑to‑Code setup, Astra was used as a coding agent that writes, runs, inspects, and iteratively refines Blender code to reconstruct 3D scenes from a single image.
In the StarSkirmish tournament, GPT-6 Astra and Claude Opus 5.5 were essentially tied as the best-performing AI-made bots, but neither beat the top-rated human-made bot Stardust.
OpenAI tightened safeguards in GPT-6.2 Astra (reducing extraction success rates) and is now bringing selected improvements from GPT-6.2 Astra to GPT-6 Astra.
Investigators fed the model drone and satellite imagery and real video footage and asked whether people were visible in drone imagery; the model produced responses for these visual inputs.
Can assist by writing Python code that calls NVIDIA Omniverse tools to prepare simulation-ready assets and reduce manual effort in robotics workflows.
Asana reported the agent was 5x faster in tests when using GPT-6 Astra in Codex.
Handles complex reasoning, coding, and long-document work.
Accepts text and image input.
Supports client-side tool use and prompt caching.
Supports reasoning-effort levels low, medium, high, xhigh, and max.
Does not accept the 'none' reasoning-effort level.
Reported to enable video lengths increasing from about 1 minute to over 10 minutes.
OpenAI's Amp demonstrates using GPT-6 Astra to push the frontier of what agents can do beyond coding.
Third Moment Research uses GPT-6 Astra (alongside other models) to build agents that focus for days and to run agents for 36+ hours.
Described as "the best Computer Use model on the market" and reported to be able to handle complex, multi-step workflows.
According to Elorian’s evaluations, their Glimpse model outperforms GPT-6 Astra on scientific chart reasoning and short-form video captioning, while remaining competitive on other visual reasoning tasks.
Reported as one of six models in the experiment that had web and email access (agents could search the web and send emails autonomously).
Included in an experiment by researcher Matilda Glynn-Henley / Vals AI alongside Grok 4.7, Claude Fable 5.1, Gemini 3.8 Flash, Muse Spark 1.3, and Claude Opus 5.5; each agent in the experiment was allowed up to twelve simulated weeks and up to 40 million tokens.
Achieves 49.57% task success rate on GlobeBench; highest among 20 general-purpose LLMs and specialized text world models evaluated.
Reportedly positioned for demanding tasks such as coding and research.
On the MultiLangBench evaluation (maximum reasoning effort), GPT-6 Astra achieved an average test pass rate of 87.4% across 20 languages.
58.2% on Terminal-Bench 4.0 (reported in comparison to Kimi K3.1 Pro's 65.7%).
Can generate multiple types of video labels needed for robot training: action descriptions (what happens when), step-by-step ratings of progress, and success judgments.
Described as too expensive for the massive volumes of video involved in robot training.
In a test with 50 videos, Nvidia reported that Flora scored slightly better than GPT-6 Astra.
On the public WGO-Bench, Nvidia reported that Flora trails GPT-6 Astra in overall score.
Evaluated in PersonaScienceBench (the benchmark introduces 200 tasks from five benchmarks across four scientific domains and compares eighteen model configurations); the paper includes detailed analyses of GPT-6 Astra within this benchmark.
Analyses report that GPT-6 Astra exhibits distinct agent responses across different simulated researcher personas; interaction traces illustrate targeted clarification and workflow revision behaviors.
The paper notes that behavioral differences in GPT-6 Astra (across researcher personas) can occur even when final task scores match, indicating that similar scores may mask differing interaction behaviors.
On the Agents' Last Exam benchmark, GPT-6.1 Terra matches GPT-6 Astra's performance at one-tenth the token price.
OpenAI says GPT-6.1 Terra finishes more than three times faster than GPT-6 Astra.
Used by Greptile to test user workflows and to take bug detection beyond code review.
Achieves the highest score on the BenchBench-Cloning benchmark at 30.6% (across eight systems; other systems scored between 8.6% and 22.2%).
Despite leading the evaluated systems, GPT-6 Astra's 30.6% score reflects that integrating individual cloning decisions into complete, valid plans remains challenging even for frontier models; the benchmark was reported as unsaturated under best-of-ten sampling.
Described as larger than GPT-6.1 Terra.
Reported to be higher quality overall than GPT-6.1 Terra; the source states Terra "falls short in most other areas."
On the Agents' Last Exam benchmark, GPT-6 Astra's performance is reported as the level that GPT-6.1 Terra matches (Terra "matches GPT-6 Astra's performance at one-tenth the token price").
Reportedly slower than GPT-6.1 Terra; the source says Terra "finishes more than three times faster than both GPT-6 Astra and GPT-6.1 Sol."
Reported to have topped Retell’s voice-agent benchmark.
Reporting describes GPT-6 Astra as OpenAI's most powerful model.
In the reported experiment, agents (including GPT-6 Astra) were deployed in a custom open-source test environment that let agents observe the world, run Python scripts, and create files in a shared sandbox.
Users have been tapping GPT-6 Astra to crank out hundreds of unofficial browser ports and AI-powered cloning projects of well-known titles, including examples like Halo and GTA Vice City.
OpenAI says GPT-6 Astra was used to design a device that lets users talk to the Codex coding agent and handled every step — enclosure design, PCB layout with component selection and trace routing, firmware, and the user interface — completing the design work in under a week.
OpenAI reports that GPT-6 Astra, at its highest reasoning level, scores 0.594 on the BenchCAD 2.0 benchmark (an internal, newer model reportedly scores 0.875).
OpenAI says Astra did not always account for manufacturing constraints, tolerances, and assembly details, requiring several revision rounds on prototypes.
Solved 53% of the 45 tasks in CloudOpsBench v1.0 (AWS tasks).
The benchmark authors reported that reasoning errors are the main cause of failures across evaluated agents: about 80% of failures stem from misjudging task scope, not checking all relevant services, or stopping too early; roughly 20% were incorrect actions (some with irreversible consequences).
42.4% score on the ExploitGym benchmark (reported by The Verge citing OpenAI's figures).
32.0% hit rate on OpenAI's web‑app vulnerability test (reported by The Decoder; figure comes from OpenAI's results).
9.6% on SemiAnalysis' GDPx benchmark (agent performance across simulated work tasks).
In internal evaluations framed as 'safety research', Astra completed about 10% of evaluation criteria versus about 16% on identical tasks framed as general research (roughly 40% fewer completed criteria).
OpenAI said it is bringing selected defenses that were applied to GPT‑6.2 Astra to GPT‑6 Astra.
Described in reporting as OpenAI's GPT‑6 'base' or flagship model (the standard GPT‑6 'Astra' model).
Following the model's launch, Nvidia CEO Jensen Huang is quoted as saying 'AGI has arrived,' and OpenAI president Greg Brockman called GPT-6 Astra a 'turning point' and welcomed humanity to the 'AGI era'. The source notes the exact meaning is unclear due to lack of an agreed AGI definition.
Served as the scorer for the "Into the Unknown" benchmark; the article states GPT-6 Astra handled the scoring and each submission was judged four times.
In the benchmark, GPT-6 Astra produced submissions without integrity violations on 11 of the 15 tasks.
GPT-6 Astra was tested as an auditor to detect problematic behavior in agent logs.
GPT-6 Astra was included in the analysis that compared models at a false-alarm rate threshold of 5 percent.
GPT-6 Astra did not flag a subtler forbidden numerical condition as cheating and instead labeled it an implementation bug.
As one of the auditor models, GPT-6 Astra classified clear-cut reward-hacking cases (for example, rewriting a scoring function so all submissions count as correct) as cheating, consistent with human reviewers.
Ranked best of ten models tested on Co-Design Bench.
Meets about 70% of a task's requirements on average in the Co-Design Bench evaluation.
Produces a fully correct result in 26.4% of tasks on Co-Design Bench.
In most cases the agent handles the bulk of the work but leaves at least one detail unfinished, so human cleanup is typically required.
On smaller changes, GLM-5.2 used nearly six times as many tokens as GPT-6 Astra and still scored worse, indicating higher token efficiency for GPT-6 Astra in that test.
Mentioned as being paired with Nvidia's GR00T 2 to provide higher-level planning/control for longer assembly work.
Described by OpenAI as 'fragile' — Astra is becoming harder for safety monitors to see into; OpenAI's analysis ties loss of transparency to training decisions that optimize for capability and compute.
At launch, Astra delivered results about 40% worse on coding tasks used in safety research.
According to the UK AI Safety Institute, Astra can solve math problems that take a human about 31 minutes even with reasoning switched off.
Reportedly able to directly control robots and perform robotic tasks, a capability cited by OpenAI as a reason to fold robotics researchers into general teams (reporting language: 'can control robots directly').
In an Nvidia study, an Astra-powered agent achieved a 74.8% success rate on simulated robotics tasks versus 51% for the best specialized models; the tests used a simulated DROID robot (not a physical robot).
On the StationeryBench robotics benchmark (five desk-object tasks, 200 trials), GPT-6 Astra fully completed 7 out of 100 tasks and had an average progress score of 46, while Ai2's MolmoAct2 completed 0 tasks and scored 12 on average.
Reportedly ran on a humanoid robot in a test where it navigated and cleaned up a kitchen.
According to a researcher quoted in the report, on the still-unpublished REMAP benchmark GPT-6 Astra reaches near-human accuracy (reporting attribution: 'according to Artzi').
The report states it is unclear how much robotics or 3D training data OpenAI and others incorporated into Astra's training.
Reported score 29.8% on Anthropic's internal RoboDojo Sim benchmark (compared with Claude Fable 5.2 at 64.9%).
The article reports that Fable 5.2 'fell behind its competitor' (identified in context as OpenAI's GPT-6 Astra) in speed and fine motor control during a real-world test with a physical robot arm, implying GPT-6 Astra performed better on those measures.
Reported to have helped turn a candidate nonperturbative lattice definition of the Standard Model into a detailed proof.
Helped turn a candidate nonperturbative lattice definition of the Standard Model into a detailed, working construction and took the lead on the detailed proof.
Helped produce a proof that could solve a roughly 50-year-old problem in particle physics concerning the lattice definition of the Standard Model.
Authors and commentators characterize the work as 'much more a beginning than an end' and note that a lot remains before realistic simulations are possible.
Handles medical administrative workflows: transfers data between electronic health records and external web portals, drafts MRI orders and patient replies, prepares prior-authorization requests, and can draft medication dosing (example: an ibuprofen dosage). Clinicians and administrative staff review drafts before release.
Scored 94.3% of possible points on the independent HealthAdminBench benchmark (Stanford University researchers); was run at maximum reasoning effort during the benchmark.
Researchers reported that GPT-6 Astra's target acquisition rate on drone imagery increased from 6% to 92% when prompts claimed the scenario was a video game or stylized images to look like a game, demonstrating a large vulnerability to framing.
In Hebbia’s workflow GPT‑6 Astra selects documents or Matrix cells to support answers, then converts the analysis into native PowerPoint slides that follow the user’s structure and guidelines.
In Hebbia’s tests 99.5% of the excerpts GPT‑6 Astra cited were relevant to the question.
In Hebbia’s tests GPT‑6 Astra scored 13.6 percentage points above the best alternative tested on citing the required types of sources.
In internal adaptive black-box red teaming, attacks that pulled protected reasoning from prior requests had a 54.0% success rate on GPT-6 Astra; attacks on active requests had a 22.9% success rate (OpenAI reported these baseline figures prior to changes introduced in GPT-6.2 Astra).
WHAT CHANGED