DeepSeek V4.1-Flash Claims #1 Open Model on Code Arena WebDev at 1670 ELO
Arena.ai posted a Code Arena WebDev update on October 18 that installed DeepSeek-V4.1-Flash-20261017 as the top-ranked open model on the leaderboard.
The score is 1670 ELO. That puts it at #1 among open models and #12 overall on the Code Arena WebDev leaderboard, which covers agentic and front-end coding tasks. It displaced Kimi K3 (Max), the previous open-model leader.
The cost difference between the two is approximately 10x. Arena.ai described the result as shifting the “performance-cost Pareto frontier for open models in coding workflows.”
Context
DeepSeek shipped the V4.1-Flash-1017 checkpoint on October 17. That release made the Search and Fetch APIs publicly available; native multimodal image input had shipped with the original V4.1 Flash on September 10. The Code Arena WebDev ranking reflects the same checkpoint, evaluated on Arena.ai’s coding and agentic task suite.
Prior benchmark positions for V4.1 Flash:
| Benchmark | Score |
|---|---|
| LiveBench Agentic Coding (Sep 2026) | 77.3 — #1 overall, open or closed |
| DeepSWE (V4.1-Flash-1017) | 77.0 |
| LiveBench Overall (Sep 2026) | 81.1 |
| Code Arena WebDev ELO (Oct 18) | 1670 — #1 open model |
The agentic coding leadership on LiveBench was already established before the October checkpoint. The Code Arena result is additional evidence across a different evaluation surface — Arena.ai’s human preference-based coding battles rather than automated task resolution.
What Code Arena Measures
Code Arena WebDev pairs models against each other on front-end and agentic coding tasks, with human raters judging the output. ELO is calculated from head-to-head win rates. A 1670 ELO among open models is a meaningful lead; Kimi K3 Max, the prior leader, requires substantially higher running costs to match or exceed.
Open-Weight Implications
V4.1 Flash is distributed as open weights under an MIT licence. Code Arena WebDev #1 on an open-weight model at a fraction of closed-model costs is the combination that matters for developers building production coding agents. The closed frontier — Claude Opus 5.5, Fable 5.1, GPT-6 Astra — occupies the top 11 spots overall. V4.1 Flash at #12 is the nearest open challenger.
Prior to the V4.1-Flash October checkpoint, the leading open model for coding agentic work was also V4.1 Flash on LiveBench. This result extends that lead to a second independent evaluation.