Model comparison
DeepSeek V4 Pro vs GPT-6 Astra
GPT-6 Astra is the stronger model overall, scoring 70.8 to 54.3 on the Noometry Index. DeepSeek V4 Pro costs 20× less per token, which makes it the better buy when GPT-6 Astra's lead doesn't matter for your workload.
Last verified . 41 shared benchmarks.
DeepSeekSummary
- They share 41 benchmarks with published results for both. DeepSeek V4 Pro scores higher in 2 categories and GPT-6 Astra in 7 categories; 6 gaps are clear of the uncertainty.
- The widest gap is in math, where GPT-6 Astra leads 93.5 to 64.8.
- The biggest single-benchmark swing is FrontierMath Tier 4: 26.8% for DeepSeek V4 Pro and 97.6% for GPT-6 Astra.
- DeepSeek V4 Pro is cheaper at $0.66 / $1.98 per million input/output tokens, against $10 / $50 for GPT-6 Astra.
- GPT-6 Astra accepts more context: 1.05M tokens versus 1M.
- DeepSeek V4 Pro has downloadable open weights; the other is API-only.
Side by side
| DeepSeek V4 Pro | GPT-6 Astra | |
|---|---|---|
| Provider | DeepSeek | OpenAI |
| Noometry Index | 54.3 | 70.8 |
| Released | 2026-04-24 | 2026-09-03 |
| Weights | Open | Proprietary |
| Context window | 1M | 1.05M |
| Max output | 393K | 128K |
| Input $ / M tokens | $0.66 | $10 |
| Output $ / M tokens | $1.98 | $50 |
| Results tracked | 48 | 56 |
Sponsored placements are available on pages like this one. Advertise on Noometry
Category by category
Coding GPT-6 Astra leads
DeepSeek V4 Pro: 52.4 (#34), GPT-6 Astra: 73.7 (#2)
| Benchmark | DeepSeek V4 Pro | GPT-6 Astra |
|---|---|---|
| FrontierCode | 28.6% | 53.3% |
| LMArena WebDev | 1582 | 1786 |
| SciCode | 51% | 56.5% |
| WeirdML | 66.2% | 93.6% |
| LMArena Coding | 1470 | 1487 |
| ALE-Bench | 1,403 | 2,951 |
| SWE-bench Verified | 77.6% | — |
| DeepSWE | — | 74.1% |
| FrontierSWE | — | 65.5% |
| GSO | — | 79.4% |
| MirrorCode | — | 46.7% |
Agentic & Tool Use GPT-6 Astra leads
DeepSeek V4 Pro: 32.8 (#58), GPT-6 Astra: 52.9 (#3)
| Benchmark | DeepSeek V4 Pro | GPT-6 Astra |
|---|---|---|
| APEX-Agents | 47.3% | 64.7% |
| Vending-Bench 2 | 3,285 | 15,515 |
| Remote Labor Index | — | 20.8% |
| BALROG | — | 68.3% |
| GDP.pdf | — | 34.2% |
Reasoning GPT-6 Astra leads
DeepSeek V4 Pro: 56.5 (#24), GPT-6 Astra: 85.1 (#1)
| Benchmark | DeepSeek V4 Pro | GPT-6 Astra |
|---|---|---|
| ARC-AGI-2 | 61.3% | 95% |
| NYT Connections (extended) | 91.3% | 98.1% |
| ARC-AGI-1 | 90.5% | 98.5% |
| CritPt | 18% | 31.7% |
| Chess Puzzles | 47% | 72% |
| LMArena Hard Prompts | 1461 | 1462 |
| Mystery Game Puzzles | 43% | 84% |
| DTBench | 93.9% | 97.3% |
| LMCA | 45.5% | 64.4% |
| Epoch Capabilities Index | 155.31 | 166.45 |
| Kagi LLM Benchmark | 53.5% | — |
| EBR-Bench | — | 76.2% |
| Surface Evolver Bench | 40% | — |
| Bench to the Future 3 | — | 0.14 |
| ForecastBench | 56.1 | — |
Math GPT-6 Astra leads
DeepSeek V4 Pro: 64.8 (#30), GPT-6 Astra: 93.5 (#2)
| Benchmark | DeepSeek V4 Pro | GPT-6 Astra |
|---|---|---|
| FrontierMath (Tiers 1-3) | 64.6% | 93.7% |
| FrontierMath Tier 4 | 26.8% | 97.6% |
| OTIS Mock AIME 2024-2025 | 98.6% | 100% |
| ProofBench | 50% | 99% |
| LMArena Math | 1455 | 1465 |
| MathArena Final-Answer Competitions | 76.6% | — |
| FrontierMath Erdős | — | 2.9% |
Knowledge GPT-6 Astra leads
DeepSeek V4 Pro: 59.5 (#31), GPT-6 Astra: 75.3 (#1)
| Benchmark | DeepSeek V4 Pro | GPT-6 Astra |
|---|---|---|
| GPQA Diamond | 91.7% | 95.8% |
| SimpleQA Verified | 52.9% | 75.6% |
| Vectara Hallucination Rate | 8.6% | 8.7% |
| LMArena Expert | 1464 | 1483 |
| Humanity's Last Exam | — | 54.8% |
Multimodal Not comparable
DeepSeek V4 Pro: —, GPT-6 Astra: 55.0 (#3)
| Benchmark | DeepSeek V4 Pro | GPT-6 Astra |
|---|---|---|
| LMArena Vision | — | 1281 |
| Blueprint-Bench 2 | — | 49.7% |
| Furniture Assembly | — | 80% |
| LMArena Document | — | 1468 |
Multilingual Too close to call
DeepSeek V4 Pro: 54.4 (#45), GPT-6 Astra: 53.7 (#61)
| Benchmark | DeepSeek V4 Pro | GPT-6 Astra |
|---|---|---|
| LMArena Non-English | 1439 | 1430 |
| LMArena Chinese | 1486 | 1484 |
| LMArena French | 1472 | 1456 |
| LMArena German | 1458 | 1440 |
| LMArena Japanese | 1445 | 1379 |
| LMArena Korean | 1447 | 1426 |
| LMArena Russian | 1453 | 1436 |
| LMArena Spanish | 1458 | 1407 |
Instruction Following Too close to call
DeepSeek V4 Pro: 76.1 (#47), GPT-6 Astra: 76.3 (#44)
| Benchmark | DeepSeek V4 Pro | GPT-6 Astra |
|---|---|---|
| LMArena Instruction Following | 1448 | 1450 |
Long Context Too close to call
DeepSeek V4 Pro: 45.0 (#51), GPT-6 Astra: 44.5 (#62)
| Benchmark | DeepSeek V4 Pro | GPT-6 Astra |
|---|---|---|
| LMArena Longer Query | 1458 | 1456 |
| CL-bench Life | 13.5% | — |
Writing & Preference GPT-6 Astra leads
DeepSeek V4 Pro: 65.5 (#46), GPT-6 Astra: 75.3 (#7)
| Benchmark | DeepSeek V4 Pro | GPT-6 Astra |
|---|---|---|
| LMArena Text | 1451 | 1441 |
| LMArena Creative Writing | 1446 | 1418 |
| EQ-Bench Creative Writing | 1553 | 2173 |
| LMArena Multi-Turn | 1467 | 1448 |
| EQ-Bench 4 | 1166 | — |
Frequently asked questions
Is DeepSeek V4 Pro better than GPT-6 Astra?
GPT-6 Astra is the stronger model overall, scoring 70.8 to 54.3 on the Noometry Index. DeepSeek V4 Pro costs 20× less per token, which makes it the better buy when GPT-6 Astra's lead doesn't matter for your workload.
Which is cheaper, DeepSeek V4 Pro or GPT-6 Astra?
DeepSeek V4 Pro is cheaper. It lists at $0.66 per million input tokens and $1.98 per million output tokens; GPT-6 Astra lists at $10 and $50.
Is DeepSeek V4 Pro or GPT-6 Astra better for coding?
GPT-6 Astra scores higher on coding benchmarks: 73.7 versus 52.4 in the Noometry coding category.
Which has the bigger context window?
GPT-6 Astra does, with 1.05M tokens against 1M.
How many benchmarks do DeepSeek V4 Pro and GPT-6 Astra share?
41 benchmarks have published results for both models. DeepSeek V4 Pro has 48 scored results on Noometry and GPT-6 Astra has 56.