DeepSeek releases open AI model that codes for under $0.20 a task, according to its own tests

A 3D model of the Forbidden City, a racing game playable in the browser, a recreation of Portal 2 in the Godot game engine. These are the demos DeepSeek chose on October 17 to present the new version of its fast model, DeepSeek-V4.1-Flash-1017. The Chinese company's message: a model that can carry out a software project from start to finish on its own, for a few cents.

The model is available on DeepSeek's website, app and API. Its weights can be downloaded from Hugging Face and ModelScope under the MIT license, one of the most permissive: commercial use, modification and self-hosting are allowed. The model card describes it as a "minor update" to V4.1-Flash, with a context window of one million tokens.

Shorter, faster, cheaper: how good are these claims?

DeepSeek touts a "Pareto improvement": better results at a lower cost. The model reportedly produces shorter answers and finishes its tasks faster. According to The Decoder, it scores 77% on the DeepSWE coding benchmark for a little under $0.20 per task. The previous version scored 74.2% for about $0.50. In DeepSeek's chart, rival models cost from $2 to more than $10 per task.

One caveat: all of these figures come from DeepSeek. At the time of the announcement, no independent evaluation had been published. And the table on the Hugging Face model card is more nuanced than the announcement. The new model improves on its predecessor across all eleven benchmarks. But against competitors, it leads on only a few: DeepSWE (77 versus 74 for Anthropic's Claude Opus 5), Automation-Bench, Agents' Last Exam and Apex Bench. Elsewhere, Claude Opus 5 almost always keeps the advantage, sometimes by a wide margin: 51.8 versus 42.3 on Terminal-Bench 4.0, and 56.3 versus 42.6 on Humanity's Last Exam without tools.

An open model, but not for a desktop PC

Open does not mean lightweight. According to Hugging Face, the model has 763 billion parameters in total. DeepSeek does not publish the hardware configuration required. The model card points to the serving software vLLM and SGLang, and to community-compressed versions for llama.cpp, Ollama or LM Studio. As of publication, no inference provider on Hugging Face yet offered the model.

For a European company, the MIT license nonetheless has a concrete benefit: it allows the model to be hosted on its own servers or with a provider of its choice, instead of sending data to the API of a Chinese company. Agent capabilities should still be tested on your own projects before any production use.

A search engine opened to developers

DeepSeek is also opening its internal search engine to developers, through two interfaces called Search and Fetch. According to the company, the engine handles hundreds of millions of queries per day. The price is $0.75 per 1,000 calls, per interface. For DeepSeek API users, these calls are free in the DeepSeek Harness, Claude Code and Codex tools until October 31 at 4:00 p.m. UTC.

This release comes in a busy week for open models: Meta unveiled its own while acknowledging that it still trails two Chinese rivals. A question hangs over what comes next, though. According to The Decoder, the Chinese government is reportedly considering restricting the release of open models, notably those ranking among the top ten. Will DeepSeek's next versions still be released openly?