New DeepSeek model promises better coding at a fraction of the price

DeepSeek released the model DeepSeek-V4.1-Flash-1017 on October 17, 2026. It arrives about five weeks after its predecessor, which came out on September 10. The model is available through DeepSeek's website, app and API (the access developers use to build the model into their own programs).

According to Crypto Briefing, the model is open source under an MIT license, and API prices are unchanged from the previous version. The weights, meaning the learned numbers that make up the model itself, have been published on Hugging Face and ModelScope so others can download them.

Better and cheaper, according to DeepSeek

DeepSeek itself calls the update a Pareto improvement in the balance between performance and cost. Roughly speaking, that means you get more for your money without giving up anything else. According to the company, the model solves tasks faster and gives shorter answers than its predecessor.

On the DeepSWE coding test, the model scores 77.0% at the high setting, according to DeepSeek's own figures, at roughly USD 0.20 per task. At the low setting, it scores over 72% at roughly USD 0.10 per task. Its predecessor scored 74.2% at roughly USD 0.50 per task.

Comparison with major competitors

In DeepSeek's chart, GPT-6 Astra, Gemini 3.8 Flash and Claude Opus 5 land at roughly USD 2 to over USD 10 per task with lower accuracy than the new model. Crypto Briefing reports that Claude Opus 5 scored 74.0% on DeepSWE v1.1.

According to OfficeChai, the model beats Claude Opus 5 on Agents' Last Exam (35.4 versus 28.6) and Automation-Bench (60.5 versus 50.3). It trails on Humanity's Last Exam (roughly 42.6-44.0 versus 56.3), NL2Repo-Bench and ProgramBench. The outlets also report that it beats Kimi K3, GLM 5.3 and GPT 5.6-Sol on most agent and coding tests.

Caveats

The figures come primarily from DeepSeek itself, and according to The Decoder, independent tests had not yet been published. The cost figures also differ slightly between sources: One cites roughly USD 0.15-0.20 at the lowest setting, another roughly USD 0.10. In addition, benchmarks do not necessarily reflect real-world work environments or code quality.

DeepSeek's Search and Fetch API costs USD 0.75 per 1,000 calls, but is free for users of DeepSeek's API for a limited period. The sources disagree on whether this applies until October 31 or November 1, 2026.