DeepSeek V4.1 Flash 1017: 77% on DeepSWE at 20 Cents a Task

DeepSeek’s newest small model scores 77.0 percent on the DeepSWE coding benchmark at an average cost of just under $0.20 per task, according to the company’s own chart, while larger rivals in the same chart cost between $2 and more than $10 per task. For engineering teams paying for coding agents by the task, the question is no longer which model is smartest but whether a top-tier score is worth paying 10 times as much or more.

The Number

As reported by The Decoder, DeepSeek released DeepSeek-V4.1-Flash-1017 on its website, app and API, and published the weights on Hugging Face and ModelScope. In DeepSeek’s cost chart for DeepSWE, the model reaches 77.0 percent at just under $0.20 per task. In its low setting it still clears 72 percent for about $0.10. Its direct predecessor, V4.1-Flash, scored 74.2 percent at around $0.50 per run.

The comparison models in the chart, including GPT-6 Astra, Gemini 3.8 Flash and Claude Opus 5, land between $2 and over $10 per task, and none of them match the smaller DeepSeek model’s accuracy in that chart. In DeepSeek’s wider benchmark set, the update also beats Claude Opus 5, GPT 5.6-Sol, Kimi K3 and GLM 5.3 in several coding and agent categories.

How the Math Works

The Decoder ties the cost drop to shorter outputs. DeepSeek says the new model runs faster and produces shorter outputs than its predecessor, and calls the update a Pareto improvement in the tradeoff between performance and cost. A model that reaches the same answer in fewer tokens costs less per task even at an unchanged per-token price, which is why per-task cost, not price per million tokens, is the number to watch for agent workloads.

The quality gain comes from post-training. DeepSeek uses reinforcement learning with self-feedback (RLSF), in which the model uses its own confidence scores or self-generated critiques as the reward signal, removing the need for human evaluators or a separate reward model. The company says this improves end-to-end project work, from game programming to 3D modeling and design. The model was tuned for agent harnesses including DeepSeek Harness (v0.2.1-rc.1 or later recommended), Tencent CodeBuddy and WorkBuddy.

What It Signals

DeepSeek is also opening its in-house search engine to developers through new Search and Fetch APIs, priced at $0.75 per 1,000 calls per endpoint, with setup covered in the DeepSeek API docs. API users can use them free through server-side tools in DeepSeek Harness, Claude Code and Codex until October 31, 2026. Read together, a cheap model plus cheap retrieval looks like a bid for the agent loop, where a task makes many model and search calls.

Two caveats limit how far the numbers travel. Every figure above is DeepSeek’s own; independent benchmarks have not been published yet, and The Decoder notes that developer impressions of code quality can differ sharply from benchmark results. And the Chinese government reportedly plans to tighten rules on publishing open-weight models, which adds a policy risk to building deeply on Chinese weights. This continues the pattern Stack Archive tracked in DeepSeek’s permanent pricing model: DeepSeek competes on cost per unit of work, not on topping leaderboards.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.