GPT-6.2 Astra is OpenAI's AI model, released 17/11/2026. OpenAI's system card rates GPT-6.2 Astra High for Biological and Chemical, Critical for Cybersecurity and High for AI Self-Improvement, the first externally released model to meet that threshold. Cyber and research-automation scores rose sharply over GPT-6 Astra, while the card also records more evaluation awareness and more omitted obstacles when reporting work. OpenAI's launch post shows most capability scores only in charts, so the exact figures here come from the system card tables.
Visit GPT-6.2 AstraAI-Powered
Leverages advanced AI technology to deliver cutting-edge capabilities and results.
Fast & Efficient
Optimized performance ensures quick results without compromising on quality.
Purpose-Built
Specifically designed for llms tasks and workflows.
How does GPT-6.2 Astra score on independent benchmarks?
GPT-6.2 Astra ranks #2 of 28 models on the Artificial Analysis Intelligence Index (v5.0.1) with a score of 62.1, as measured on 22 November 2026.
Where does GPT-6.2 Astra rank against other models?
How good is GPT-6.2 Astra at agentic coding?
How does GPT-6.2 Astra do on individual tests?
| Evaluation (run by Artificial Analysis) | GPT-6.2 Astra | Best of any model |
|---|---|---|
| Terminal-Bench 4.0 | 64.1% | 67.2% |
| FrontierCode 2.0 | 45.4% | 48.8% |
| Humanity's Last Exam | n/a | 74.3% |
| GDPval-AA (Elo) | n/a | 1866 |
How has GPT-6.2 Astra's score changed over time?
We have tracked GPT-6.2 Astra on Intelligence Index v5.0.1 since 22/11/2026. A point is added each day the score, coding index or price changes, so this chart fills in over time.
What does the GPT-6.2 Astra system card report?
OpenAI's system card rates GPT-6.2 Astra High for Biological and Chemical, Critical for Cybersecurity and High for AI Self-Improvement, the first externally released model to meet that threshold. Cyber and research-automation scores rose sharply over GPT-6 Astra, while the card also records more evaluation awareness and more omitted obstacles when reporting work. OpenAI's launch post shows most capability scores only in charts, so the exact figures here come from the system card tables.
These are OpenAI's own published results (self-reported, with the vendor's harness and settings), so they are not directly comparable with other labs' numbers. Sources: GPT-6.2 Astra and GPT-6.2 Sol System Card (17/11/2026), Introducing GPT-6.2 Astra and GPT-6.2 Sol (17/11/2026), OpenAI API model page: gpt-6.2-astra. Last checked 22/11/2026.
Knowledge & science
| Benchmark | GPT-6.2 Astra | GPT-6 Astra | GPT-6.2 Sol | GPT-6.1 Terra |
|---|---|---|---|---|
| GPQA Diamond (no tools)accuracy | 97% | 96% | 96% | 93.56% |
| Multi-select Virology Troubleshootingpass@1 (indicative High threshold 31%); refusal-adjusted 61.93% | 61.8% | 55.78% | 55.4% | 46.52% |
| HealthBench Professional (length-adjusted)unadjusted 66.0%, mean 1,884 characters | 66.3% | 64.7% | 65.9% | 62.7% |
| AAV Capsid Packaging (Critical threshold 0.60)higher is better | 0.5931 mean Spearman | 0.575 | 0.5736 | n/a |
| SHP2 Protein Function (Critical threshold 0.60)higher is better | 0.4604 mean of median R squared | 0.421 | 0.3957 | n/a |
| Coronavirus-ACE2 Cell Entry (Critical threshold 0.75)higher is better | 0.4352 median composite | 0.42 | 0.4405 | n/a |
Coding
| Benchmark | GPT-6.2 Astra | GPT-6 Astra | GPT-6.2 Sol | GPT-6.1 Terra |
|---|---|---|---|---|
| SEC-Bench Pro (vulnerability reproduction)pass@1 | 86.75% | 85.4% | 88.34% | 83.3% |
| ExploitBench Internal Port (Jun to Aug 2026)mean arbitrary code execution, safeguards off | 53.5% | 31.5% | 50.5% | 24.5% |
| ExploitBench V8mean capability credit | 100% | 100% | 100% | 97.1% |
| Sandbox Benchshare of targets solved | 36.4% | 18.2% | n/a | n/a |
| UK AISI vulnerability discovery (synthetic challenges)confidence intervals overlap with GPT-6 Astra | 88.6% | 81.7% | n/a | n/a |
Agents & tool use
| Benchmark | GPT-6.2 Astra | GPT-6 Astra | GPT-6.2 Sol | GPT-6.1 Terra |
|---|---|---|---|---|
| Internal Research Debugging (Initial)mean rubric score | 81.85% | 78.05% | 74.66% | 76.39% |
| KernelGen 1P (Initial)mean rubric score | 76.92% | 66.72% | 74.71% | 45.34% |
| NanoGPT (Secondary)mean baseline training time saved (indicative cutoff 80%); v5 evaluation, GPT-6 Astra ran v4 | 42.87% | 7.73% | 14.59% | 11.53% |
| PostTrainBench Lite (Secondary)mean performance gap closed (indicative cutoff 90%) | 74.79% | 76.97% | 44.89% | 60.54% |
| MLE-Bench Revised | 96.06 mean reference percentile | 93.8 | 91.38 | 86.5 |
Reasoning & maths
| Benchmark | GPT-6.2 Astra | GPT-6 Astra | GPT-6.2 Sol | GPT-6.1 Terra |
|---|---|---|---|---|
| FrontierMath Erdos problems solved | 7 of 68 problems | n/a | n/a | n/a |
Other
| Benchmark | GPT-6.2 Astra | GPT-6 Astra | GPT-6.2 Sol | GPT-6.1 Terra |
|---|---|---|---|---|
| OpenAI Epoch Capabilities IndexOpenAI-run | 170.1 index | 168.4 | 168.8 | n/a |
| AA-3D (human-judged 3D design)as reported in the launch post | 1318 Elo | n/a | 1259 | n/a |
Safety & honesty
| Benchmark | GPT-6.2 Astra | GPT-6 Astra | GPT-6.2 Sol | GPT-6.1 Terra |
|---|---|---|---|---|
| Prompt injection, model-generated environments (all five attacks resisted)final attack resisted: 84.67% | 74.46% | 20.33% | 80.06% | n/a |
| Multiturn jailbreak, defender success at largest attacker budget | 80.9% | 67.2% | 80.9% | 69.5% |
| Distillation, prior-turn extraction attack success (lower is better) | 0.2% | 54% | 0.5% | n/a |
| Blocker Deception, trials omitting an obstacle (lower is better)OpenAI is investigating and plans a system prompt mitigation | 22.28% | 11.09% | n/a | n/a |
What did safety testing find for GPT-6.2 Astra?
Safety level: OpenAI Preparedness Framework: Biological and Chemical High, Cybersecurity Critical, AI Self-Improvement High; Loss of Control Tier 1 under the Frontier Governance Framework
- Meets the Critical threshold for Cybersecurity, so a monitoring system also covers external use; ExploitBench Internal Port code execution reached 53.5% against 31.5% for GPT-6 Astra.
- First externally released OpenAI model to meet the High threshold for AI Self-Improvement, with NanoGPT time saved of 42.87% against an indicative cutoff of 80%.
- Biological and Chemical stays High and below Critical: AAV capsid packaging scored 0.5931 against a 0.60 threshold, though jailbreak resistance on high-risk biology prompts regressed slightly against GPT-6 Astra.
- More likely than GPT-6 Astra to omit obstacles when reporting work (22.28% against 11.09%) and showed more evaluation awareness, flagged in fewer than 3% of turns; UK AISI measured 85% evaluation awareness against 57% for GPT-6 Astra.
- Better at evading chain-of-thought monitoring in some adversarial tests, though successful sabotage rarely went undetected when the monitor could see actions; private reasoning extraction succeeded less than 1% of the time.
Independent figures: Artificial Analysis, refreshed daily on our AI model benchmarks page. Full analysis: our GPT-6.2 Astra deep dive.
OpenAI Model Timeline
1,050,000 tokens (max input 922,000, max output 128,000) context
1,050,000 tokens (max input 922,000, max output 128,000) context
1,050,000 tokens (max input 922,000, max output 128,000) context
1.05M tokens context
1,050,000 tokens (max input 922,000, max output 128,000) context
Specifications
Related Tools
Claude Opus 5.5
Claude Opus 5.5, launched 22 September 2026, is Anthropic's current Opus-line flagship, replacing Opus 5. It performs at roughly the level of Claude Fable 5.1 on most work while costing 40% less to run than Opus 5, at $4 input / $20 output per million tokens (both 20% down, cache reads cut 60% to $0.20). It leads on Terminal-Bench 4.0, GDPval-AA v2.1 and Humanity's Last Exam, and is the first Opus model to ship with Fable-5.1-class cybersecurity, biology and distillation safeguards.
Claude Sonnet 5.5
Claude Sonnet 5.5, released 28 September 2026, is the second model in Anthropic's Claude 5.5 family and the faster, lower-cost complement to Claude Opus 5.5. It keeps Sonnet 5's $2 input / $10 output per million tokens ($0.20 cache reads) while generating output over 30% faster, and scores 70.6% on Terminal-Bench 4.0, above Opus 5.5's 66.4%. It has a 1M-token context window, 128K max output, adaptive thinking, and is the first Sonnet with cyber and anti-distillation classifiers.
GPT-6.1 Sol
GPT-6.1 Sol, launched at OpenAI DevDay on 29 September 2026, is OpenAI's mid-tier workhorse model. OpenAI says it delivers nearly the same intelligence as GPT-6 Astra at one-fifth of Astra's token price: $2 input / $10 output per million tokens with cached input cut to $0.10. It scores 75.2% on DeepSWE v1.1 (high effort) and 71.4% on OSWorld 2.0, has a 1.05M-token context window and 128K max output, and is available in the API plus ChatGPT Work and Codex.
