Open Source · OCTOBER 17, 2026
DeepSeek's V4.1-Flash-1017 Ships at the Same Price, Takes DeepSWE to 77.0, and Opens a Live-Web Search API
A silent same-API-ID update lifts Terminal-Bench 4.0 from 31.2 to 42.3 and puts DeepSeek ahead of Claude Opus 5 on DeepSWE v1.1, while a new first-party Search endpoint prices web retrieval at $0.75 per 1,000 calls.
DeepSeek shipped V4.1-Flash-1017 on October 17, routing the new checkpoint behind the existing deepseek-flash API ID on web, app, and server with no change to calling method or pricing. The silent swap is the point. A same-price, higher-score drop into the same endpoint is the clearest expression yet of the strategy DeepSeek telegraphed last month: hold the rate card, move the frontier underneath it.
The numbers do the arguing. Against the September 10 Flash release, Terminal-Bench 4.0 climbs from 31.2 to 42.3, DeepSWE v1.1 from 74.2 to 77.0, and ProgramBench from 20.3 to 29.1. Terminal-Bench runs were conducted under Claude Code; DeepSWE under Mini-SWE-V2. On DeepSeek's own comparison table, as read by MarkTechPost, Flash-1017's 77.0 sits above Claude Opus 5 at 74.0 and GPT-5.6 Sol at 73.0. HLE without tools is the honest counterweight: Flash-1017 posts 42.6, while Opus 5 holds 56.3. Agentic coding is where the lead is; raw knowledge isn't.
The second shoe is the Search API. POST /v1/tools/search and POST /v1/tools/fetch on api.deepseek.com give agents a first-party live-web layer at $0.75 per 1,000 successful calls (5 yuan, per BigGo Finance), capped at 5 QPS per account, with results carrying a 0–3 rerank score, a 0–4 authority tier, and site:, lang:, after: and before: operators. Server-side calls through DeepSeek Harness, Claude Code, or Codex are free until 16:00 UTC on October 31; after that, the economics are the message.
In a 56-trial pilot with 14 pass criteria, Kingy.ai ran Flash-1017 against GPT-6.1 Sol and Claude Sonnet 5.5. Total cost: $0.0157 for Flash-1017, $0.0577 for GPT-6.1 Sol, $0.0653 for Claude Sonnet 5.5. One pilot isn't a benchmark, but it's the shape of the argument.
What's new structurally is the integration of model and retrieval under one meter. Builders running competitive monitoring, prospect research, or outreach-prep loops have been paying two subscriptions: one for inference, one for search. DeepSeek's bet is that collapsing those into a single sub-dollar unit, on open weights distributed via Hugging Face and ModelScope, prices the stack below what a hyperscaler can match without cannibalizing its own search partners. It's the same move Flash made on inference pricing, extended one layer up the agent loop, and consistent with the architectural cost story DeepSeek began telling in September.
November 1 is the dated test. If $0.75 per 1,000 holds under scaled agent traffic after the free window closes, the cost floor for autonomous web work resets for everyone else.