DeepSeek Put Its Own Search Engine on the Price List. The Model Update Was the Smaller Half

DeepSeek shipped two things on October 17. One is DeepSeek-V4.1-Flash-1017, a post-training checkpoint of the Flash model it released in September; its own Hugging Face card calls it "a minor update." The other is a pair of API endpoints that sell access to DeepSeek's in-house web search index at $0.75 per thousand calls. The checkpoint took the headlines. The endpoints are the larger story.

What the checkpoint changes

The model is served as deepseek-flash, and the weights are on Hugging Face under an MIT licence at 763B parameters in the published safetensors. The architecture carries over from September: a multimodal mixture-of-experts with a 1M-token context and up to 384K output tokens, 8B active parameters per token in prefill and 16B in decode. What changed is post-training. DeepSeek says it trained the checkpoint with reinforcement learning from self-feedback — the model's own confidence scores and self-generated critiques as the reward signal, rather than human raters or a separate reward model — and tuned the result for end-to-end project work inside its own agent harness.

The numbers below are DeepSeek's own. No independent evaluation exists yet.

The shape is consistent. Large gains over its own predecessor everywhere, a lead over Claude Opus 5 across the agentic and automation suites, and a clear deficit on raw reasoning and on the longer software-engineering benchmarks. Against the other open-weight contenders DeepSeek claims 77.0 on DeepSWE to Kimi K3's 67.5 and GLM-5.3's 66.9, with the smallest total parameter count of the three.

The cost figures are what the release is actually arguing. DeepSeek puts the 77.0 run at just under $0.20 per task, and a lower reasoning-effort setting above 72% at roughly $0.10. September's checkpoint scored 74.2 at about $0.50 per run. In DeepSeek's own comparison chart, the models in that accuracy band cost from about $2 to over $10 per task. Token pricing is $0.30 per million cache-miss input tokens and $1.20 per million output at peak hours, and half that off-peak.

Selling the index

Alongside the model, DeepSeek opened POST /v1/tools/search and POST /v1/tools/fetch. Search returns ranked results from what the company describes as its in-house search engine, with a 0–3 rerank score blending relevance, authority and freshness, a 0–4 authority tier for the source, short snippets written for a user interface and longer summaries written for a model, and the operators a developer would expect: site:, lang:, after:, before:. Fetch takes a single URL and returns the page as Markdown.

Labs have bundled web search into their chat products for two years. Metering the crawl itself, by the call, to any developer — including one who never touches the model — is a different kind of product. It also discloses something about DeepSeek that a checkpoint does not: the company operates a general web index at a scale it is willing to put behind a paid endpoint with a published rate limit. For the first two weeks after launch, searches routed through Claude Code's built-in web search tool or Codex's web_search are exempt from the fee. That is less a promotion than a demonstration. The cheapest way to get a Western coding agent's retrieval layer running on a Chinese index is to make it free for a fortnight.

The stack, not the model

Read together, the two announcements describe a company assembling the entire agent stack and pricing every layer beneath the model close to zero. There is the harness — DeepSeek Harness, now in developer preview, with the checkpoint explicitly tuned for it and for Tencent's CodeBuddy. There is the model, at a tenth to a fiftieth of the per-task cost of the frontier it claims to match on agentic work. And now there is retrieval, sold by the call.

That is a competitive position built on unit economics rather than capability leadership, and the benchmark table above concedes the capability gap openly: on reasoning without tools, DeepSeek is fourteen points behind. The bet is that for an agent running thousands of tool calls across a long task, the cost per task decides adoption and the last few points of reasoning do not.

Two cautions belong with the numbers. All of them are self-reported, on a checkpoint DeepSeek itself calls minor, by a lab that has tuned the model to its own harness — and the agentic benchmark where it trails Opus 5 by the widest margin, Terminal-Bench 4.0, is the one DeepSeek ran inside Claude Code rather than its own scaffold. Whether the gains survive a move to someone else's harness is the first thing independent evaluation should test. The second caution is political: reports this week indicate Beijing is preparing tighter rules on publishing open-weight models, with a model's capability ranking among the criteria for review. The MIT licence on these weights is a current fact, not a guaranteed one.

Sources: DeepSeek-V4.1-Flash-1017 announcement and API docs — DeepSeek · Search and Fetch API guide — DeepSeek · Models and pricing — DeepSeek · DeepSeek-V4.1-Flash-1017 model card — Hugging Face · Deepseek V4.1 Flash 1017 beats larger rivals in agentic tasks while once again massively cutting per-task costs — The Decoder · DeepSeek Just Released V4.1-Flash-1017, Scores 77.0 on DeepSWE, Opens Search API — MarkTechPost