GPT-6 Astra
| Released | 2026-09-04 |
|---|---|
| Price per 1M | $10.00 in · $50.00 out |
| Overall score | 65.4 · rank 4 |
| Context window | 1,050,000 tokens |
| Provider | OpenAI |
| Weights | Proprietary |
Scores
| Dimension | Score | Rank | Benchmarks |
|---|---|---|---|
| Overall: Balanced | 65.4 | 4 | 57 |
| Overall: Coding | 65.2 | 4 | 44 |
| Overall: Reasoning | 67.9 | 4 | 20 |
| Overall: Agentic work | 65.1 | 4 | 41 |
| Overall: Frontend | 64.2 | 6 | 16 |
| Coding | 65.1 | 7 | 12 |
| Agents | 64.0 | 4 | 16 |
| Reasoning | 69.5 | 2 | 13 |
| Math | 68.3 | 3 | 4 |
| Knowledge | 63.5 | 10 | 3 |
| Frontend | 64.5 | 5 | 3 |
| General | 65.1 | 8 | 5 |
| Multimodal | 60.2 | 19 | 1 |
| Value vs price | +8.7 | 41 |
Buzz
| Posts on X | About 195 a day |
|---|---|
| Mood | 12% positive, 24% negative |
| Hacker News | 29 mentions this week |
| Checked | 2026-11-11 |
- Unverified rumours 😃 Physical AI anxiety and Huawei Ascend pretraining: Sources close to DeepSeek note that OpenAI's GPT 6 Astra hitting 95% accuracy on a robotic control task (versus 40% for Claude Fable 5.1) triggered internal concern that scaling text and code LLMs alone wi…
- TOWARD WEIGHT-SPACE PROGRAMMING MIT CSAIL researchers recently had a frontier model, GPT-6 Astra, fix another AI by editing its weights directly. No training. They call it weight-space programming. Could a model running on my own Mac do that kind of brain surgery? Test 1 The…
- Which model has the best overall intelligence? - Gemini 4 Carbon - Fable 5.2 - Opus 5.5 - GPT-6.2 Sol - GPT-6 Astra
Configurations
| Configuration | Balanced | Benchmarks |
|---|---|---|
| Unspecified setting | 67.8 ±1.4 | 16 |
| High | 64.1 ±1.6 | 12 |
| Low | 62.7 ±2.1 | 8 |
| Max | 64.5 ±0.9 | 39 |
| Medium | 64.0 ±2.1 | 8 |
| Non-reasoning | 60.6 ±4.0 | 6 |
| Extra High | 65.3 ±2.4 | 10 |
Evidence
Coding 65.1 ±2.0
| Benchmark | Result | Source | Date |
|---|---|---|---|
| FrontierSWE | 65.5% | Epoch AI | 2026-11-11 |
| DeepSWE | 74.1% | Epoch AI | 2026-11-11 |
| FrontierCode | 53.3% | Epoch AI | 2026-11-11 |
| Code Migration v1.1 | 80.6% | Vals AI | 2026-11-09 |
| LiveBench 2.0 Coding | 87.5% | LiveBench | 2026-11-06 |
| WeirdML v3 | 42.2% | Epoch AI | 2026-11-11 |
| MirrorCode | 60.0% | Epoch AI | 2026-10-29 |
| GSO | 79.4% | Epoch AI | 2026-11-11 |
| IOI | 100.0% | Vals AI | 2026-10-29 |
| SWE-bench Multimodal | 43.3% | Vals AI | 2026-11-09 |
| SciCode | 56.5% | Epoch AI | 2026-11-11 |
| ProgramBench | 5.5% | Vals AI | 2026-11-09 |
Agents 64.0 ±1.2
| Benchmark | Result | Source | Date |
|---|---|---|---|
| Terminal-Bench 4 | 60.3% | Terminal-Bench | 2026-09-03 |
| SRE Bench | 56.9% | Vals AI | 2026-11-05 |
| Terminal-Bench 4.0 (Vals run) | 59.6% | Vals AI | 2026-11-09 |
| Terminal-Bench-Science | 62.9% | Vals AI | 2026-11-09 |
| APEX-Agents | 64.7% | Epoch AI | 2026-11-11 |
| Remote Labor Index | 20.8% | Epoch AI | 2026-11-11 |
| LiveBench 2.0 Agentic | 75.6% | LiveBench | 2026-11-06 |
| LiveBench 2.0 Browser Use | 91.3% | LiveBench | 2026-11-06 |
| Terminal-Bench 2.1 (Vals run) | 87.3% | Vals AI | 2026-10-01 |
| EMB v1.1 (Excel financial modeling) | 72.7% | Vals AI | 2026-11-09 |
| LiveBench 2.0 Research | 90.6% | LiveBench | 2026-11-06 |
| Tax Agent Bench | 20.7% | Vals AI | 2026-11-09 |
| Finance Agent v2.1 | 60.4% | Vals AI | 2026-11-09 |
| BALROG | 68.3% | Epoch AI | 2026-11-11 |
| Vending-Bench 2 | $15,515 | Epoch AI | 2026-11-11 |
| HLAB-Vals 1.1 (legal agent) | 7.1% | Vals AI | 2026-11-09 |
| CyberBench v1.1 | 41.1% | Vals AI | 2026-11-05 |
Reasoning 69.5 ±1.3
| Benchmark | Result | Source | Date |
|---|---|---|---|
| ARC-AGI-2 | 95.0% | Epoch AI | 2026-11-11 |
| ARC-AGI-2 (ARC Prize verified) | 95.0% | ARC Prize | 2026-11-11 |
| Extended NYT Connections | 98.4% | Lech Mazur | 2026-11-10 |
| MysteryMechanism | 53.2% | Vals AI | 2026-11-09 |
| GPQA Diamond | 95.8% | Epoch AI | 2026-08-30 |
| CritPt | 31.7% | Epoch AI | 2026-11-11 |
| ARC-AGI-1 (ARC Prize verified) | 98.5% | ARC Prize | 2026-11-11 |
| ARC-AGI-1 | 98.5% | Epoch AI | 2026-11-11 |
| SimpleBench | 83.6% | SimpleBench | 2026-09-07 |
| BioMysteryBench | 79.3% | Vals AI | 2026-11-09 |
| ARC-AGI-3 | 100.0% | ARC Prize | 2026-11-11 |
| Chess Puzzles | 74.0% | Epoch AI | 2026-10-13 |
| LiveBench 2.0 Reasoning | 92.7% | LiveBench | 2026-11-06 |
| Mystery Game Puzzles | 84.0% | Epoch AI | 2026-08-30 |
| LiveBench 2.0 Data Analysis | 83.0% | LiveBench | 2026-11-06 |
Math 68.3 ±1.2
| Benchmark | Result | Source | Date |
|---|---|---|---|
| FrontierMath Tiers 1-3 | 93.7% | Epoch AI | 2026-08-30 |
| FrontierMath Tier 4 | 97.6% | Epoch AI | 2026-08-30 |
| OTIS Mock AIME 2024-2025 | 100.0% | Epoch AI | 2026-08-30 |
| ProofBench v1.1 | 99.0% | Vals AI | 2026-11-09 |
| ProofBench | 99.0% | Epoch AI | 2026-11-11 |
Knowledge 63.5 ±3.6
| Benchmark | Result | Source | Date |
|---|---|---|---|
| SimpleQA Verified | 75.6% | Epoch AI | 2026-08-30 |
| Humanity's Last Exam | 54.8% | Epoch AI | 2026-11-11 |
| Legal Research Bench | 39.4% | Vals AI | 2026-11-09 |
Frontend 64.5 ±2.7
| Benchmark | Result | Source | Date |
|---|---|---|---|
| LiveBench 2.0 Design | 80.0% | LiveBench | 2026-11-06 |
| Vibe Code Bench v1.1 | 89.6% | Vals AI | 2026-11-09 |
| LMArena WebDev | 1,767 | LMArena | 2026-11-09 |
General 65.1 ±4.6
| Benchmark | Result | Source | Date |
|---|---|---|---|
| LLM Debate Benchmark | 1,604 | Lech Mazur | 2026-10-26 |
| EQ-Bench Creative Writing v3 | 2,177 | EQ-Bench | 2026-11-11 |
| LMArena Text | 1,474 | LMArena | 2026-11-09 |
| Creative Story-Writing (Lech Mazur) | 91.0% | Lech Mazur | 2026-10-23 |
| EQ-Bench Longform Creative Writing | 82.8% | EQ-Bench | 2026-11-11 |
Multimodal 60.2 ±4.1
| Benchmark | Result | Source | Date |
|---|---|---|---|
| LMArena Vision | 1,283 | LMArena | 2026-11-09 |
Results not used in the score
| Benchmark | Result | Source | Why not used |
|---|---|---|---|
| ARC-AGI-1 | 97.5% | Epoch AI | A stronger configuration result was used |
| ARC-AGI-1 | 86.0% | Epoch AI | A stronger configuration result was used |
| ARC-AGI-1 | 96.5% | Epoch AI | A stronger configuration result was used |
| ARC-AGI-1 | 97.5% | Epoch AI | A stronger configuration result was used |
| ARC-AGI-1 | 98.5% | Epoch AI | A stronger configuration result was used |
| ARC-AGI-1 (ARC Prize verified) | 97.5% | ARC Prize | A stronger configuration result was used |
| ARC-AGI-1 (ARC Prize verified) | 86.0% | ARC Prize | A stronger configuration result was used |
| ARC-AGI-1 (ARC Prize verified) | 96.5% | ARC Prize | A stronger configuration result was used |
| ARC-AGI-1 (ARC Prize verified) | 97.5% | ARC Prize | A stronger configuration result was used |
| ARC-AGI-1 (ARC Prize verified) | 98.5% | ARC Prize | A stronger configuration result was used |
| ARC-AGI-2 | 59.6% | Epoch AI | A stronger configuration result was used |
| ARC-AGI-2 | 85.4% | Epoch AI | A stronger configuration result was used |
| ARC-AGI-2 | 92.1% | Epoch AI | A stronger configuration result was used |
| ARC-AGI-2 | 92.1% | Epoch AI | A stronger configuration result was used |
| ARC-AGI-2 | 93.3% | Epoch AI | A stronger configuration result was used |
| ARC-AGI-2 (ARC Prize verified) | 59.6% | ARC Prize | A stronger configuration result was used |
| ARC-AGI-2 (ARC Prize verified) | 85.4% | ARC Prize | A stronger configuration result was used |
| ARC-AGI-2 (ARC Prize verified) | 92.1% | ARC Prize | A stronger configuration result was used |
| ARC-AGI-2 (ARC Prize verified) | 92.1% | ARC Prize | A stronger configuration result was used |
| ARC-AGI-2 (ARC Prize verified) | 93.3% | ARC Prize | A stronger configuration result was used |
| ARC-AGI-3 | 62.7% | ARC Prize | A stronger configuration result was used |
| ARC-AGI-3 | 98.6% | ARC Prize | A stronger configuration result was used |
| ARC-AGI-3 | 35.2% | ARC Prize | A stronger configuration result was used |
| ARC-AGI-3 | 96.7% | ARC Prize | A stronger configuration result was used |
| ARC-AGI-3 | 17.5% | ARC Prize | A stronger configuration result was used |
| ARC-AGI-3 | 98.0% | ARC Prize | A stronger configuration result was used |
| ARC-AGI-3 | 38.6% | ARC Prize | A stronger configuration result was used |
| ARC-AGI-3 | 98.4% | ARC Prize | A stronger configuration result was used |
| ARC-AGI-3 | 54.8% | ARC Prize | A stronger configuration result was used |
| ARC-AGI-3 | 59.3% | ARC Prize | A stronger configuration result was used |
| ARC-AGI-3 | 98.4% | ARC Prize | A stronger configuration result was used |
| CritPt | 19.7% | Epoch AI | A stronger configuration result was used |
| CritPt | 26.3% | Epoch AI | A stronger configuration result was used |
| CritPt | 29.1% | Epoch AI | A stronger configuration result was used |
| CritPt | 28.9% | Epoch AI | A stronger configuration result was used |
| CritPt | 31.4% | Epoch AI | A stronger configuration result was used |
| DeepSWE | 73.2% | Epoch AI | A stronger configuration result was used |
| DeepSWE | 67.0% | Epoch AI | A stronger configuration result was used |
| DeepSWE | 72.8% | Epoch AI | A stronger configuration result was used |
| DeepSWE | 73.2% | Epoch AI | A stronger configuration result was used |
| Extended NYT Connections | 97.7% | Lech Mazur | A stronger configuration result was used |
| Extended NYT Connections | 98.1% | Lech Mazur | A stronger configuration result was used |
| FrontierMath Tier 4 | 82.9% | Epoch AI | A stronger configuration result was used |
| FrontierMath Tier 4 | 87.8% | Epoch AI | A stronger configuration result was used |
| FrontierMath Tier 4 | 97.6% | Epoch AI | A stronger configuration result was used |
| FrontierMath Tier 4 | 97.6% | Epoch AI | A stronger configuration result was used |
| FrontierMath Tier 4 | 97.6% | Epoch AI | A stronger configuration result was used |
| MirrorCode | 46.7% | Epoch AI | A stronger configuration result was used |
| SciCode | 53.5% | Epoch AI | A stronger configuration result was used |
| SciCode | 54.1% | Epoch AI | A stronger configuration result was used |
| SciCode | 54.2% | Epoch AI | A stronger configuration result was used |
| SciCode | 55.4% | Epoch AI | A stronger configuration result was used |
| SciCode | 55.7% | Epoch AI | A stronger configuration result was used |
| Terminal-Bench 4 | 50.9% | Terminal-Bench | A stronger configuration result was used |
| Terminal-Bench 4 | 55.2% | Terminal-Bench | A stronger configuration result was used |
| Terminal-Bench 4 | 60.0% | Terminal-Bench | A stronger configuration result was used |
| Terminal-Bench 4 | 59.7% | Terminal-Bench | A stronger configuration result was used |
Names matched to this model
gpt-6-astra