Up to four models
GPT-6 Astra vs Gemini 3.8 Flash
On the index (GPT-4 = 100), GPT-6 Astra scores 199 and Gemini 3.8 Flash 175. 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
199182–215
WeightsClosed
ReleasedSep 2026
Price in / out$10 / $50
Context1.05M
Core tests taken7 of 9
Gemini 3.8 Flash175159–191
WeightsClosed
ReleasedSep 2026
Price in / out$0.82 / $4.13
Context1.05M
Core tests taken7 of 9
GPT-6 Astra leads Gemini 3.8 Flash by 24 points, ahead on 6 of 7 shared core tests, at 12× the price.
Test by testthird-party results
GPT-6 AstraGemini 3.8 Flash
Gemini 3.8 FlashEvery result as a table
| Test | GPT-6 Astra | Gemini 3.8 Flash |
|---|---|---|
| Reasoning & knowledge | ||
| GPQA Diamond (Epoch AI) | 95.8% | 95.4% |
| Humanity's Last Exam | 54.8% | 44.5% |
| SimpleQA Verified | 75.6% | 69.7% |
| Maths | ||
| FrontierMath (tiers 1-3) | 93.7% | 68.4% |
| AIME-style maths (OTIS mock) | 100.0% | 98.9% |
| ProofBench v1.1 | 99.0% | 48.0% |
| Chess puzzles | 74.0% | 58.0% |
| FrontierMath (tier 4) | 97.6% | 21.9% |
| Coding | ||
| Terminal-Bench 4.0 | 59.6% | 19.7% |
| Code Migration | 67.7% | 36.5% |
| CyberBench v1.1 | 41.1% | 43.8% |
| IOI | 100.0% | 56.9% |
| ProgramBench | 5.5% | 1.0% |
| Vibe Code Bench 1-100 | 27.6% | 18.8% |
| Vibe Code Bench v1.1 | 89.6% | 78.6% |
| DeepSWE | 74.1% | 73.8% |
| FrontierSWE | 65.5% | 19.6% |
| WeirdML | 93.3% | 84.8% |
| Agents | ||
| APEX-Agents | 64.7% | 64.3% |
| BioMysteryBench | 79.3% | 62.2% |
| SRE Bench | 56.9% | 11.1% |
| Terminal-Bench-Science | 62.9% | 8.6% |
| DTBench | 97.3% | 95.7% |
| Furniture assembly | 80.0% | 31.7% |
| LMCA | 64.4% | 52.9% |
| Mystery game puzzles | 84.0% | 47.0% |
| Professional work | ||
| Finance Agent (v2) | 53.5% | 61.4% |
| EMB | 71.7% | 72.2% |
| HLAB-Vals-1.1 | 7.1% | 12.1% |
| Legal Research Bench | 39.4% | 38.9% |
| MedCode | 48.5% | 48.1% |
| MedScribe | 87.9% | 84.5% |
| SAGE | 46.4% | 35.1% |
| Tax Agent Bench | 20.7% | 32.1% |
| Novel problems | ||
| ARC-AGI-2 | 95.0% | 89.2% |
| ARC-AGI-1 | 98.5% | 98.5% |
| ARC-AGI-3 | 62.7% | 10.4% |
| MysteryMechanism | 53.1% | 36.5% |
Tests at least two of these models have taken. Bold rows are the core tests.
