Up to four models
GPT-6 Astra vs DeepSeek-V4-Pro
On the index (GPT-4 = 100), GPT-6 Astra scores 193 and DeepSeek-V4-Pro 163. Side by side on the same third-party tests, with price and context. The link in your address bar shares this exact comparison.
GPT-6 Astra
193178–209
WeightsClosed
ReleasedSep 2026
Price in / out$10 / $50
Context1.05M
Core tests taken7 of 9
163147–178
WeightsOpen
ReleasedAug 2026
Price in / out$0.11 / $4.44
Context1.05M
Core tests taken6 of 9
GPT-6 Astra leads DeepSeek-V4-Pro by 31 points, ahead on 6 of 6 shared core tests, at 17× the price.
Test by testthird-party results
GPT-6 Astra
DeepSeek-V4-Pro
Every result as a table
| Test | GPT-6 Astra | DeepSeek-V4-Pro |
|---|---|---|
| Reasoning & knowledge | ||
| GPQA Diamond (Epoch AI) | 95.8% | 91.7% |
| LiveBench Data Analysis | 83.0% | 79.2% |
| LiveBench IF | 75.6% | 67.7% |
| LiveBench Language | 89.4% | 82.1% |
| LiveBench Reasoning | 92.7% | 85.8% |
| SimpleQA Verified | 75.6% | 52.9% |
| Maths | ||
| FrontierMath (tiers 1-3) | 93.7% | 64.6% |
| AIME-style maths (OTIS mock) | 100.0% | 98.6% |
| LiveBench Mathematics | 96.8% | 95.1% |
| ProofBench v1.1 | 99.0% | 50.0% |
| Chess puzzles | 74.0% | 46.0% |
| FrontierMath (tier 4) | 97.6% | 26.8% |
| Coding | ||
| Terminal-Bench 4.0 | 59.6% | 14.1% |
| LiveBench Agentic Coding | 57.3% | 54.9% |
| LiveBench Coding | 80.4% | 77.2% |
| Code Migration | 67.7% | 41.5% |
| CyberBench v1.1 | 41.1% | 72.9% |
| IOI | 100.0% | 51.6% |
| ProgramBench | 5.5% | 0.0% |
| Vibe Code Bench 1-100 | 27.6% | 17.5% |
| Vibe Code Bench v1.1 | 89.6% | 82.3% |
| MirrorCode | 60.0% | 33.3% |
| WeirdML | 93.3% | 66.2% |
| Agents | ||
| APEX-Agents | 64.7% | 47.3% |
| SRE Bench | 56.9% | 1.9% |
| Terminal-Bench-Science | 62.9% | 4.3% |
| DTBench | 97.3% | 93.9% |
| LMCA | 64.4% | 45.5% |
| Mystery game puzzles | 84.0% | 43.0% |
| Professional work | ||
| Finance Agent (v2) | 53.5% | 50.4% |
| EMB | 71.7% | 52.8% |
| HLAB-Vals-1.1 | 7.1% | 10.8% |
| Legal Research Bench | 39.4% | 40.9% |
| MedCode | 48.5% | 42.5% |
| MedScribe | 87.9% | 80.2% |
| Tax Agent Bench | 20.7% | 26.8% |
| Novel problems | ||
| ARC-AGI-2 | 95.0% | 61.3% |
| ARC-AGI-1 | 98.5% | 90.5% |
Tests at least two of these models have taken. Bold rows are the core tests.