Side-by-side benchmark comparison across agentic, coding, multimodal, knowledge, reasoning, and math workflows.
GPT-5.4
82
Winner · 0/8 categoriesHolo3-35B-A3B
~78
1/8 categoriesGPT-5.4· Holo3-35B-A3B
Pick GPT-5.4 if you want the stronger benchmark profile. Holo3-35B-A3B only becomes the better choice if agentic is the priority or you want the cheaper token bill.
GPT-5.4 is clearly ahead on the aggregate, 82 to 78. The gap is large enough that you do not need to squint at the spreadsheet to see the difference.
GPT-5.4 is also the more expensive model on tokens at $2.50 input / $15.00 output per 1M tokens, versus $0.25 input / $1.80 output per 1M tokens for Holo3-35B-A3B. That is roughly 8.3x on output cost alone. GPT-5.4 is the reasoning model in the pair, while Holo3-35B-A3B is not. That usually helps on harder chain-of-thought-heavy tests, but it can also mean more latency and more token spend in real use. GPT-5.4 gives you the larger context window at 1.05M, compared with 64K for Holo3-35B-A3B.
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 | Holo3-35B-A3B |
|---|---|---|
| AgenticHolo3-35B-A3B wins | ||
| Terminal-Bench 2.0 | 75.1% | — |
| BrowseComp | 82.7% | — |
| OSWorld-Verified | 75% | 77.8% |
| MCP Atlas | 67.2% | — |
| Toolathlon | 54.6% | — |
| tau2-bench | 98.9% | — |
| Coding | ||
| HumanEval | 95% | — |
| SWE-bench Verified | 84% | — |
| LiveCodeBench | 84% | — |
| SWE-bench Pro | 57.7% | — |
| React Native Evals | 82.6% | — |
| Multimodal & Grounded | ||
| MMMU-Pro | 81.2% | — |
| OfficeQA Pro | 96% | — |
| MMMU-Pro w/ Python | 81.5% | — |
| OmniDocBench 1.5 | 0.1090 | — |
| Reasoning | ||
| MuSR | 94% | — |
| BBH | 97% | — |
| MRCRv2 | 97% | — |
| MRCR v2 64K-128K | 86% | — |
| MRCR v2 128K-256K | 79.3% | — |
| Graphwalks BFS 128K | 93.1% | — |
| Graphwalks Parents 128K | 89.8% | — |
| ARC-AGI-2 | 73.3% | — |
| Knowledge | ||
| MMLU | 99% | — |
| GPQA | 92.8% | — |
| SuperGPQA | 96% | — |
| MMLU-Pro | 93% | — |
| HLE | 48% | — |
| FrontierScience | 91% | — |
| HLE w/o tools | 39.8% | — |
| SimpleQA | 97% | — |
| Instruction Following | ||
| IFEval | 96% | — |
| Multilingual | ||
| MGSM | 96% | — |
| MMLU-ProX | 94% | — |
| Mathematics | ||
| AIME 2023 | 99% | — |
| AIME 2024 | 99% | — |
| AIME 2025 | 99% | — |
| HMMT Feb 2023 | 96% | — |
| HMMT Feb 2024 | 98% | — |
| HMMT Feb 2025 | 97% | — |
| BRUMO 2025 | 97% | — |
| MATH-500 | 99% | — |
GPT-5.4 is ahead overall, 82 to 78. The biggest single separator in this matchup is OSWorld-Verified, where the scores are 75% and 77.8%.
Holo3-35B-A3B has the edge for agentic tasks in this comparison, averaging 77.8 versus 77. Inside this category, OSWorld-Verified 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.