DeepSeek-V4.1-Flash-20261017 has taken the top open-model spot on a coding leaderboard. On the Code Arena WebDev leaderboard published by Arena on October 18, 2026, the build released on October 17 scored 1,670 points, ranking first among open models and 12th overall. The model it displaced is Kimi K3 (Max), and Arena puts its running cost at roughly a tenth of that rival's.
This No. 1 is not a vendor benchmark
Two kinds of scores need separating. DeepSeek's launch numbers, such as DeepSWE and Terminal-Bench, are vendor-run results on fixed test sets, already covered in our DeepSeek-V4.1-Flash-1017 launch report. The Code Arena WebDev leaderboard measures something else: real users hand agentic coding and front-end development tasks to different models and rank them by preference votes, testing whose generated web pages and front-end apps people actually prefer.
The two kinds of evidence carry different weight. Vendor scores can be shaped by test-set selection and tuning direction, while preference leaderboards have no standard answer and fluctuate more, but track what developers feel day to day. That is exactly why this result merits its own story: an open model did not just look good in its own table, it also reached the top of the open rankings in third-party, blind-style voting.
Low cost is not a side note here
The model uses a Mixture-of-Experts design with 552 billion backbone parameters plus 196 billion Engram parameters, activating only about 8 billion parameters during prefill and about 16 billion during decoding, with a 1-million-token context window. Weights are published on Hugging Face under the MIT license. Teams that prefer not to self-host can use the API at peak pricing of $0.30 per million input tokens and $1.20 per million output tokens.
Arena describes the result as shifting the performance-cost frontier for open models in coding workflows. For engineering teams, the practical meaning is more specific than the ranking: when a model placed 12th overall costs a tenth of the leading open rival, routing a share of frequent, low-risk coding tasks to it stops being an experiment and becomes an option worth costing properly. The precondition is a small trial on real tasks from your own repositories, not swapping the default model on the strength of a leaderboard.
The other side of 12th overall
There is a ceiling to note: 11 models still sit above it, and WebDev covers only the front-end and agentic coding slice, saying nothing about general reasoning, long documents, or multimodal ability. Preference leaderboards also move with the voting sample, so today's first place is not guaranteed next month.
For teams choosing a model, the sound reading is to treat this top spot as a signal to trial, not a conclusion: compare completion and rework rates on one or two weeks of real tickets against your current model, then combine that with the $0.30 input price to work out cost per task. The leaderboard tells you it is worth trying; the ledger tells you whether to switch.
