Across 45 shared benchmarks, Claude Sonnet 4.5 scores higher on 24 and Qwen3.5 122B A10B on 21. The widest gap is OmniScience Non-Hallucination, where Claude Sonnet 4.5 scores 50.9 against 12.9. Qwen3.5 122B A10B is the cheaper of the two on tracked API pricing ($0.40 against $3.00 per million input tokens).
| Benchmark | Claude Sonnet 4.5 | Qwen3.5 122B A10B |
|---|---|---|
| AA Agentic Index | 50.6 | 21.3 |
| AA Intelligence | 37.4 | 32.8 |
| AA-LCR | 68.3 | 70.3 |
| AA-Omniscience | -0.1 | -41.5 |
| AI2D | 87 | 93.3 |
| Artificial Analysis Coding Index | 52.1 | 45.7 |
| browsecomp | 24.1 | 63.8 |
| browsecomp_zh | 42.4 | 69.9 |
| CC-OCR | 68.1 | 81.8 |
| coding_arena_elo | 1392 | 1358 |
| critpt | 1.1 | 0.9 |
| gdpval | 40.4 | 24.3 |
| GPQA Diamond | 83.4 | 86.6 |
| HallusionBench | 59.9 | 67.6 |
| HLE | 19.8 | 47.5 |
| HMMT 2025 | 74.6 | 90.3 |
| IFBench | 57.3 | 76.1 |
| LiveCodeBench v6 | 64 | 78.9 |
| longbench_v2 | 61.8 | 60.2 |
| MathVista | 79.8 | 87.4 |
| mmlu_redux | 95.6 | 94 |
| MMLU-Pro | 88.2 | 86.7 |
| mmmlu | 89.1 | 86.7 |
| MMMU | 79.6 | 83.9 |
| MMMU-Pro | 68.7 | 76.9 |
| multichallenge | 55.3 | 61.5 |
| OmniDocBench 1.5 | 85.8 | 89.8 |
| OmniScience Accuracy | 32.9 | 24.4 |
| OmniScience Non-Hallucination | 50.9 | 12.9 |
| OSWorld-Verified | 61.4 | 58 |
| RealWorldQA | 70.3 | 85.1 |
| scicode | 45 | 42 |
| Seal-0 | 53.4 | 44.1 |
| SimpleVQA | 57.6 | 0.6 |
| SWE-bench Verified | 82 | 72 |
| TauBench V3 - Banking | 24.5 | 15.3 |
| Terminal-Bench 2.0 | 50 | 49.4 |
| Terminal-Bench 2.1 | 55.8 | 47.6 |
| Terminal-Bench Hard | 35.6 | 31.1 |
| vectara_answer_rate | 95.6 | 99.8 |
| vectara_avg_summary_length | 127.8 | 86.4 |
| vectara_factual_consistency | 88 | 88.8 |
| vectara_hallucination_rate ↓ | 12 | 11.2 |
| τ²-Bench Telecom (AA run) | 78.1 | 93.6 |
| τ³-Bench | 19 | 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 (anthropic-official, direct), otherwise the lowest tracked offer.