DeepSeek updates its Flash model and moves the race to cost per task

DeepSeek made DeepSeek-V4.1-Flash-1017 available on October 17, 2026 on its website, in its app and via API, publishing the weights on Hugging Face and ModelScope. The model card states an MIT license for both the repository and the weights, and describes the release as a minor update to the previous V4.1-Flash: improvements in reasoning, coding, agentic tasks and visual benchmarks, with faster execution and shorter outputs. The model accepts multimodal input, image and text; the context window is 1 million tokens, and a max_tokens of at least 256K is recommended. On post-training, the company says it used reinforcement learning with self-feedback: the reward comes from the model's own confidence or from critiques it generates itself.

Here are the numbers, new version versus previous: DeepSWE v1.1 77.0 vs 74.2; Terminal-Bench 4.0 42.3 vs 31.2; HLE without tools 42.6 vs 36.8; ProgramBench 29.1 vs 20.3; Automation-Bench 60.5 vs 54.8. In the same table published by DeepSeek, Claude Opus 5 leads on HLE (56.3), Terminal-Bench 4.0 (51.8), ProgramBench (37.0) and NL2Repo-Bench (75.3); on DeepSWE v1.1 it trails at 74.0, as does GPT-5.6 Sol at 73.0. DeepSeek sums it all up as “Stronger benchmark performance, faster task completion, and shorter outputs than the previous version”, and The Decoder reports that the company presents the result as a “Pareto improvement” between performance and cost. The Decoder, again citing DeepSeek's data, puts DeepSWE at 77% accuracy for about $0.20 per task and over 72% for about $0.10 in “low” mode, versus roughly $0.50 per task for its predecessor.

Alongside the model, DeepSeek opened its Search and Fetch APIs to everyone, at $0.75 per 1,000 calls for each endpoint, free until October 31, 2026 at 16:00 UTC for API users who use them as server-side tools in DeepSeek Harness, Claude Code and Codex. The model is optimized for DeepSeek Harness (recommended version v0.2.1-rc.1 or later), Tencent CodeBuddy and WorkBuddy. Several gaps remain, though, and they determine how much can really be done with this release. Benchmarks and costs per task all come from the company, and there are no independent checks yet. The parameter count is not consistent: Hugging Face lists 763 billion in the “Model size” field, some secondary outlets speak of 552B plus 196B “Engram” parameters, one source attributes the 763B figure to the weights of the September version, and nobody states the active parameters. Per-token API pricing does not appear in the official announcement, and the figures reported by the press differ. Hardware requirements for running it locally are not given.

The part that interests me most isn't the DeepSWE score, it's the axis it's placed on. Stating $0.20 per task next to accuracy shifts the comparison from “how good is it” to “how much does it cost to get there”, and with downloadable MIT weights anyone could, in theory, redo the math. In theory: without hardware requirements and without per-token pricing, the number that makes this release interesting is also the one that, outside DeepSeek, no one has yet been able to verify.

— Olya