DeepSeek Ships V4.1-Flash-1017 as MIT Open Weights, Claims It Beats Opus 5 and GPT-5.6 Sol on Several Agent Benchmarks

DeepSeek has upgraded its V4.1 Flash model to a new “1017” snapshot, published the weights under the MIT license, kept API pricing unchanged, and opened its in-house search engine to outside developers.

DeepSeek has upgraded its V4.1 Flash model to a new snapshot called DeepSeek-V4.1-Flash-1017, according to the company’s API changelog dated Oct. 17. The new version is already live in DeepSeek’s web chat and app, and API users get it automatically under the existing model name with no price change. The Hugging Face model card calls it a “minor update,” publishes the weights under the permissive MIT license, and lists a context window of 1 million tokens.

On DeepSeek’s own benchmark table, the update scores 77.0 on DeepSWE v1.1, up from 74.2 for the previous Flash release and ahead of the 74.0 it lists for Claude Opus 5 and 73.0 for GPT-5.6 Sol. It also leads its comparison set on Automation-Bench, Agents’ Last Exam and Apex Bench. It does not win everywhere: on DeepSeek’s own chart, Opus 5 stays well ahead on Humanity’s Last Exam without tools (56.3 vs. 42.6) and Terminal-Bench 4.0 (51.8 vs. 42.3), and also leads on ProgramBench and NL2Repo-Bench. All of these are company-run results, not independent tests.

The bigger claim is cost. DeepSeek says the new snapshot finishes tasks faster with shorter outputs. The Decoder, citing a DeepSeek cost chart, reports the model reached its 77.0 DeepSWE score at just under $0.20 per task, versus roughly $0.50 per task for its predecessor; that is a reported company figure. Vercel’s AI Gateway lists the model at $0.15 per million input tokens and $0.60 per million output tokens. In the same update, DeepSeek opened its in-house search engine to developers through new Search and Fetch APIs; its description of the engine’s scale is the company’s claim.

What this means for you: Open models are now close to, and on some agent tasks claimed to be ahead of, the best closed systems at a fraction of the price. If you build with AI agents, this is worth a side-by-side test; if you just use chatbots, expect cheaper and faster AI tools to keep coming.

DeepSeek Ships V4.1-Flash-1017 as MIT Open Weights, Claims It Beats Opus 5 and GPT-5.6 Sol on Several Agent Benchmarks

Sources

One email. Only the signal.