DeepSeek released DeepSeek-V4.1-Flash-1017 on October 17, an updated checkpoint of the V4.1-Flash model it launched in September, and made it live the same day on its website, its app and its API under the existing deepseek-flash model name. The company calls the release a Pareto improvement in performance and cost: higher benchmark scores, faster task completion and shorter outputs, at unchanged API prices.
Alongside the model, DeepSeek opened two developer tools to every API user, a Search API and a Fetch API built on its own in-house search engine. For agent builders, that turns DeepSeek from a low-cost model supplier into a supplier of the web layer agents need as well, as readers of our explainer on what DeepSeek is will recognise as the next step in its open, low-price strategy.
DeepSeek-V4.1-Flash-1017 Benchmarks: Gains On All 11 Tests
All of the scores in DeepSeek’s release table are the company’s own figures, and none has been independently reproduced yet. On DeepSWE v1.1, a software-engineering benchmark, the new checkpoint scores 77.0, up from 74.2 for the previous Flash release and ahead of Claude Opus 5 at 74.0 and GPT-5.6 Sol at 73.0 in DeepSeek’s own comparison.
The largest gains are in agentic and terminal work. Terminal-Bench 4.0 rises from 31.2 to 42.3, ProgramBench from 20.3 to 29.1, Automation-Bench from 54.8 to 60.5, and Agents’ Last Exam from 31.8 to 35.4. DeepSeek says the checkpoint beats its predecessor on all 11 benchmarks it reports, by between 2.0 and 11.1 points.
The table also shows where the model does not lead. Claude Opus 5 still leads on Humanity’s Last Exam, Terminal-Bench 4.0, ProgramBench and NL2Repo-Bench in DeepSeek’s figures. The honest reading is a much stronger coding and agent model at Flash prices, not a model that has overtaken the frontier on every measure, a distinction our guide to how to read an AI model launch sets out in detail.
How The 1017 Checkpoint Was Trained And Served
DeepSeek says the update was trained with reinforcement learning with self-feedback to improve end-to-end project work, with examples ranging from building complete games to producing 3D models. It is tuned for DeepSeek’s own agent harness, DeepSeek Harness, and for Tencent’s CodeBuddy and WorkBuddy, and the company recommends DeepSeek Harness v0.2.1-rc.1 or later.
The underlying V4.1-Flash design is a multimodal mixture-of-experts model with a 1 million token context window. The model card lists 763 billion total parameters, with only part of the network active for each token, and weights released on Hugging Face and ModelScope under the MIT licence, which allows commercial use, modification and redistribution.
Because the API model name does not change, existing customers received the new checkpoint automatically. Peak API pricing in DeepSeek’s documentation is $0.30 per million cache-miss input tokens and $1.20 per million output tokens, with off-peak rates at half that level, and DeepSeek’s cost chart puts a top-setting DeepSWE task at just under $0.20, against roughly $0.50 for the previous version.
DeepSeek Search And Fetch APIs: $0.75 Per 1,000 Calls
The Search API returns ranked web results for a query, and the Fetch API returns a single page as Markdown, through endpoints on api.deepseek.com. DeepSeek says its in-house search engine serves hundreds of millions of requests a day. Each endpoint costs $0.75 per 1,000 successful requests, with empty results and failures free, and each is rate-limited to 5 queries per second per account.
Calls made through server-side tools in DeepSeek Harness, Claude Code and Codex stay free for DeepSeek API users until November 1, 2026, Beijing time, when per-request billing begins. That pricing puts DeepSeek in direct competition with dedicated search-API providers for agent workloads, at the moment agent builders are choosing the stack compared in our DeepSeek vs ChatGPT breakdown.
Why It Matters
The central entity is DeepSeek-V4.1-Flash-1017, and its significance is the combination, not any single score. An open-weight checkpoint that leads a respected coding benchmark in its maker’s own table, costs a fraction of rival models per task, and now ships with a first-party search layer changes the maths for every team building coding agents. Anthropic and OpenAI sell performance at premium prices; DeepSeek is selling good-enough-to-lead performance, plus the retrieval plumbing agents need, behind one inexpensive API key. If independent testing confirms even part of the claimed gains, pressure on frontier pricing, already the theme of the Chinese model race, intensifies again.
Background
DeepSeek, the Hangzhou-based lab backed by the quantitative fund High-Flyer, upended the AI market in early 2025 when its R1 reasoning model matched leading US models at a fraction of the reported training cost, briefly shaking US tech stocks. Its V4.1-Flash line, launched in September 2026, is the fast, cheap tier of its current range, aimed at developers who run models at scale rather than at consumers chatting in an app. The 1017 checkpoint, named for its October 17 release date, continues that pattern: incremental architecture, aggressive post-training, open weights and prices set to make switching away from closed APIs a financial argument as much as a technical one.
Source: MarkTechPost, October 17, 2026; Dealroom News, October 17, 2026
0 Comments