Mistral Large 4 1029 Hits 95.4 on LiveBench 2.0 Coding but 17.4 on Browser Use, Up 19.6 Points Overall
Mistral Large 4 1029, the open-weight build released October 29, scores 63.5 overall on LiveBench 2.0. The preview build of the same model, listed separately on the board, scores 43.9. The 19.6-point jump is uneven: coding and research moved by tens of points, browser use barely moved off the floor.
Preview vs Weights, Category by Category
| Category | Preview | 1029 weights | Change |
|---|---|---|---|
| Overall | 43.9 | 63.5 | +19.6 |
| Agentic | 8.8 | 27.5 | +18.7 |
| Design | 65.6 | 76.3 | +10.7 |
| Coding | 60.0 | 95.4 | +35.4 |
| Browser Use | 0.6 | 17.4 | +16.8 |
| Data Analysis | 76.5 | 76.6 | +0.1 |
| Research | 28.1 | 79.7 | +51.6 |
| Reasoning | 83.9 | 85.2 | +1.3 |
| Cost per successful task | $1.620 | $0.487 | -70% |
Research gained the most (+51.6), followed by coding (+35.4). Data Analysis and Reasoning, where the preview was already competitive, did not move.
Coding Is at the Frontier
A 95.4 on coding puts the open weights ahead of GPT-6 Astra (93.7), GPT-6.1 Terra (94.3) and Claude Sonnet 5.5 (94.7), and behind only Claude Fable 5.1 (97.7), Claude Opus 5.5 (96.3), GPT-6.1 Sol (96.0) and DeepSeek V4.1 Flash 1017 (96.0). At $0.487 per successful task, it costs about a tenth of what Opus 5.5 does ($4.820) on the board’s cost-adjusted metric, though Sol ($0.819) remains the benchmark for value among closed models.
Browser Use Is the Hole
Browser Use is 17.4. Among models on the current board, only Inkling (12.0), Gemma 4 31B (0.0) and Mistral’s own preview (0.6) score lower. GPT-6.1 Sol leads the category at 93.3.
The gap sits awkwardly next to Mistral’s own release claims. The weights launched with an 84.5% BrowseComp result, the best among open models. BrowseComp measures finding hard-to-locate information through web search. LiveBench’s Browser Use category tests multi-step interaction with web interfaces. Strong retrieval does not carry over to navigating forms, modals and redirects, and the 67-point spread between the two numbers is the clearest evidence of that in this release.
Agentic Remains Low
The Agentic score of 27.5 is an 18.7-point gain on the preview but still 50.4 points behind Claude Fable 5.1 (77.9). Agentic carries a 2x weight in LiveBench’s overall formula, which is why a 95.4 coding result sits alongside a 63.5 overall: the board ranks Mistral Large 4 1029 below Muse Wonder 1.0 (74.2), DeepSeek V4.1 Flash 1017 (72.9) and Kimi K3.1 Pro (71.3).
For teams choosing an open model to self-host, the profile is specific. It suits code generation, repository work and research synthesis. It is a poor fit for web-agent pipelines or long-horizon tool use.