Across 36 shared benchmarks, DeepSeek-V3 scores higher on 6 and Qwen3.5 122B A10B on 30. The widest gap is HLE, where Qwen3.5 122B A10B scores 47.5 against 5.2. DeepSeek-V3 is the cheaper of the two on tracked API pricing ($0.27 against $0.40 per million input tokens).
| Benchmark | DeepSeek-V3 | Qwen3.5 122B A10B |
|---|---|---|
| AA Agentic Index | 1.6 | 21.3 |
| AA Intelligence | 15.4 | 32.8 |
| AA-LCR | 41.3 | 70.3 |
| AA-Omniscience | -37.6 | -41.5 |
| Artificial Analysis Coding Index | 23 | 45.7 |
| C-Eval | 90.1 | 91.9 |
| critpt | 0 | 0.9 |
| gdpval | 0 | 24.3 |
| GPQA Diamond | 68.4 | 86.6 |
| GSM8K | 96.7 | 94.5 |
| HLE | 5.2 | 47.5 |
| HMMT 2025 | 27.5 | 90.3 |
| IFBench | 41 | 76.1 |
| ifeval | 87.3 | 93.4 |
| LiveCodeBench v6 | 46.9 | 78.9 |
| longbench_v2 | 48.7 | 60.2 |
| mmlu_prox | 70.5 | 82.2 |
| mmlu_redux | 90.5 | 94 |
| MMLU-Pro | 81.2 | 86.7 |
| mmmlu | 79.4 | 86.7 |
| multichallenge | 31.4 | 61.5 |
| OJBench | 24 | 39.5 |
| OmniScience Accuracy | 25.4 | 24.4 |
| OmniScience Non-Hallucination | 23.3 | 12.9 |
| scicode | 35.8 | 42 |
| supergpqa | 53.7 | 67.1 |
| SWE-bench Verified | 42 | 72 |
| TauBench V3 - Banking | 4.7 | 15.3 |
| Terminal-Bench 2.1 | 16.9 | 47.6 |
| Terminal-Bench Hard | 15.2 | 31.1 |
| vectara_answer_rate | 97.5 | 99.8 |
| vectara_avg_summary_length | 81.7 | 86.4 |
| vectara_factual_consistency | 93.9 | 88.8 |
| vectara_hallucination_rate ↓ | 6.1 | 11.2 |
| τ²-Bench Telecom (AA run) | 47.1 | 93.6 |
| τ³-Bench | 4.7 | 13.6 |
Best tracked score per model per benchmark (default configuration; source-attributed). ↓ marks lower-is-better metrics. Open either model for its full surface, provenance and pricing. Quoted rates are the price-setter row we currently track for each model — its direct or vendor-official listing where one exists (direct, direct), otherwise the lowest tracked offer.