DeepSeek Releases V4.1-Flash-1017, Claims Coding Benchmark Lead Over Claude Opus 5 at Unchanged API Prices

DeepSeek released an update to its AI model on Saturday, Oct. 17. The Hangzhou-based startup says the new version is faster, leaner and cheaper to run than the one it replaces.

The model is called DeepSeek-V4.1-Flash-1017. The company announced it in a series of posts on X, saying the update has "stronger benchmark performance, faster task completion, and shorter outputs than the previous version." It is live on DeepSeek's website, its app and its developer platform, under the existing deepseek-flash endpoint. API pricing is unchanged.

DeepSeek described it as a minor update to V4.1-Flash, which launched in early September. That makes this refresh roughly five weeks old as a successor.

The benchmark claims

The headline number is on DeepSWE v1.1, a coding test. DeepSeek's own comparison shows the new model at 77.0, up from 74.2 for the previous version. The same table lists Anthropic's Claude Opus 5 at 74.0 and OpenAI's GPT-5.6 Sol at 73.0.

DeepSeek also says it leads its table on Automation-Bench, Agents Last Exam and Apex Bench.

It does not lead everywhere. By DeepSeek's own table, Claude Opus 5 still tops Humanity's Last Exam, Terminal-Bench 4.0, ProgramBench and NL2Repo-Bench. DeepSeek says the new model improved on Humanity's Last Exam over its predecessor, but did not pass Anthropic's flagship there.

These are the company's numbers, published on X and in a technical report on Hugging Face. No independent evaluation of the new model had been published as of Saturday, Oct. 17.

Cost is the real pitch

DeepSeek calls the update "a Pareto improvement in performance and cost." Shorter outputs matter because AI usage is typically billed by the volume of text a model generates. Same answer, fewer words, lower bill, even with the price list untouched.

On per-task cost, the claims are aggressive. One summary of the release puts DeepSeek's effective cost at roughly $0.10 to $0.20 per task, under one-tenth of comparable U.S. products. Another cites figures under $0.20 to $0.25 for DeepSeek against $2 to $10 or more for competitors. Those are estimates tied to DeepSeek's own benchmark setups, not audited prices.

The weights are open-sourced under the MIT license on Hugging Face and ModelScope. Anyone can download and run them.

That is the structural contest. Bloomberg described the rapid launches as underscoring the intensifying fight between China's open-weight models and the proprietary models of U.S. rivals, which are top performers but carry far higher costs.

Search tools and agents

DeepSeek is also widening what it sells. It opened its Search and Fetch APIs to outside developers, powered by an in-house search engine the company says handles hundreds of millions of requests a day. Each endpoint costs $0.75 per 1,000 calls.

Calls made through server-side tools in DeepSeek Harness, Claude Code and Codex stay free for DeepSeek API users until per-request billing starts on Nov. 1, 2026, Beijing time. The model is tuned for DeepSeek's own agent harness and for Tencent's CodeBuddy and WorkBuddy.

The company says reinforcement learning with self-feedback sharpened its agentic and coding skills. In practice, that means asking the model to build a complete project, such as a game or a 3D model, rather than just a code snippet.

Market damage from the last release

The September V4.1-Flash release came with steep discounts. Bloomberg reported that shares of MiniMax Group, Z.ai and Alibaba Group Holding, which drives the Qwen models, plummeted on the Hong Kong stock exchange after it.

DeepSeek's money is also moving. CNBC reported that its latest funding round has grown to between $12 billion and $15 billion at a valuation of about $75 billion, ahead of a planned Shanghai IPO. DeepSeek has not commented in the material tied to this release.

The capacity problem

Cheap models only help if the servers stay up. Earlier this week DeepSeek suffered an outage of about five hours that left its servers unreachable. Bloomberg reported that some experts attributed the glitch to insufficient computing resources.

Slowdowns and service disruptions have increased in recent months, according to Bloomberg, as developers flock to the company's low prices. A price advantage means little to a developer whose production app is offline.

Anthropic and OpenAI have not publicly responded to the benchmark comparisons.

The next checkpoints are concrete. Independent testers can now run the open weights against Opus 5 and GPT-5.6 Sol on their own harnesses. Nov. 1 is when DeepSeek begins billing for Search and Fetch calls, which will show whether developers stay once the free period ends.