Side-by-side benchmark comparison across agentic, coding, multimodal, knowledge, reasoning, and math workflows.
GPT-5.4 mini
68
1/8 categoriesQwen3.5-122B-A10B
71
Winner · 5/8 categoriesGPT-5.4 mini· Qwen3.5-122B-A10B
Pick Qwen3.5-122B-A10B if you want the stronger benchmark profile. GPT-5.4 mini only becomes the better choice if agentic is the priority or you need the larger 400K context window.
Qwen3.5-122B-A10B has the cleaner overall profile here, landing at 71 versus 68. It is a real lead, but still close enough that category-level strengths matter more than the headline number.
Qwen3.5-122B-A10B's sharpest advantage is in knowledge, where it averages 81.6 against 57.4. The single biggest benchmark swing on the page is OSWorld-Verified, 72.1% to 58%. GPT-5.4 mini does hit back in agentic, so the answer changes if that is the part of the workload you care about most.
GPT-5.4 mini is also the more expensive model on tokens at $0.75 input / $4.50 output per 1M tokens, versus $0.00 input / $0.00 output per 1M tokens for Qwen3.5-122B-A10B. That is roughly Infinityx on output cost alone. GPT-5.4 mini gives you the larger context window at 400K, compared with 262K for Qwen3.5-122B-A10B.
BenchLM keeps the benchmark table and the operator tradeoffs on the same page so a better score does not hide a materially slower, pricier, or smaller-context model.
Runtime metrics show N/A when BenchLM does not have a sourced snapshot for that exact model. The scoring rules and freshness policy are documented on the methodology page.
| Benchmark | GPT-5.4 mini | Qwen3.5-122B-A10B |
|---|---|---|
| AgenticGPT-5.4 mini wins | ||
| Terminal-Bench 2.0 | 60% | 49.4% |
| OSWorld-Verified | 72.1% | 58% |
| MCP Atlas | 57.7% | — |
| Toolathlon | 42.9% | — |
| tau2-bench | 93.4% | 79.5% |
| BrowseComp | — | 63.8% |
| CodingQwen3.5-122B-A10B wins | ||
| SWE-bench Pro | 54.4% | — |
| SWE-bench Verified | — | 72% |
| LiveCodeBench | — | 78.9% |
| Multimodal & GroundedQwen3.5-122B-A10B wins | ||
| MMMU-Pro | 76.6% | 76.9% |
| MMMU-Pro w/ Python | 78% | — |
| OmniDocBench 1.5 | 0.1263 | — |
| ReasoningQwen3.5-122B-A10B wins | ||
| MRCRv2 | 40.7% | — |
| MRCR v2 64K-128K | 47.7% | — |
| MRCR v2 128K-256K | 33.6% | — |
| Graphwalks BFS 128K | 76.3% | — |
| Graphwalks Parents 128K | 71.5% | — |
| LongBench v2 | — | 60.2% |
| KnowledgeQwen3.5-122B-A10B wins | ||
| GPQA | 88% | 86.6% |
| HLE | 41.5% | — |
| HLE w/o tools | 28.2% | — |
| MMLU-Pro | — | 86.7% |
| SuperGPQA | — | 67.1% |
| Instruction FollowingQwen3.5-122B-A10B wins | ||
| IFEval | 87.4% | 93.4% |
| Multilingual | ||
| MMLU-ProX | — | 82.2% |
| Mathematics | ||
| MATH-500 | 97.4% | — |
Qwen3.5-122B-A10B is ahead overall, 71 to 68. The biggest single separator in this matchup is OSWorld-Verified, where the scores are 72.1% and 58%.
Qwen3.5-122B-A10B has the edge for knowledge tasks in this comparison, averaging 81.6 versus 57.4. Inside this category, GPQA is the benchmark that creates the most daylight between them.
Qwen3.5-122B-A10B has the edge for coding in this comparison, averaging 76.3 versus 54.4. GPT-5.4 mini stays close enough that the answer can still flip depending on your workload.
Qwen3.5-122B-A10B has the edge for reasoning in this comparison, averaging 60.2 versus 40.7. GPT-5.4 mini stays close enough that the answer can still flip depending on your workload.
GPT-5.4 mini has the edge for agentic tasks in this comparison, averaging 65.6 versus 56. Inside this category, OSWorld-Verified is the benchmark that creates the most daylight between them.
Qwen3.5-122B-A10B has the edge for multimodal and grounded tasks in this comparison, averaging 76.9 versus 76.6. Inside this category, MMMU-Pro is the benchmark that creates the most daylight between them.
Qwen3.5-122B-A10B has the edge for instruction following in this comparison, averaging 93.4 versus 87.4. Inside this category, IFEval is the benchmark that creates the most daylight between them.
Get notified when new models drop, benchmark scores change, or the leaderboard shifts. One email per week.
Free. No spam. Unsubscribe anytime. We only store derived location metadata for consent routing.