DeepSeek Ships Its Own Search API — $0.75 per 1,000 Successful Calls, Free Until November 1

Verdict: The model update is the headline; the API change is the story. On October 17 DeepSeek shipped DeepSeek-V4.1-Flash-1017 and, with far less noise, opened its in-house Search and Fetch APIs to all developers — $0.75 per 1,000 successful requests for each, failed calls don't count against billing, and everything is free until November 1, 2026. If you run agents that need grounded web results, that's the cheapest named price I've seen on a first-party search endpoint, and the free window is a real deadline, not a marketing line: per-request billing starts at 00:00 Beijing time on Nov 1. Worth wiring up this week if search-backed agents were on your roadmap.

What exactly changed at the API layer

Three things, straight from DeepSeek's official docs:

1. Search & Fetch are now standalone, per-request APIs. Search returns web results from DeepSeek's own engine; Fetch reads a page and hands back markdown. Search supports restricting by site, language and publish date. Pricing is symmetrical — $0.75 per 1,000 successful requests for both, which means an agent that fails half its lookups pays nothing for them. One sharp edge worth knowing: if you use the DeepSeek API through DeepSeek Harness, Claude Code or Codex, server-side tool calls already hit these endpoints today — those calls are free now, and they will start billing on the same Nov 1 clock.

2. The 1017 snapshot silently replaced the old models. The current model version is DeepSeek-V4.1-Flash-1017 (1M context, 384K max output). Older aliases like deepseek-v4.1-flash and deepseek-v4-flash-vision-exp are still accepted, but the models behind them are retired — your requests are now served by the 1017 snapshot at the same Flash price. No migration work needed, but if you benchmark against last month's outputs, you're no longer talking to the same model.

3. Nothing changed on token pricing. Flash stays on its existing rate card with off-peak at half of peak (peak hours remain 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday). The 1017 update claims better DeepSWE agent scores and shorter outputs, which — if it holds up in your workload — is the rare combination that lowers effective cost per task without a price cut.

The model underneath: "wish-driven programming"

Per DeepSeek's official docs, V4.1-Flash-1017 was trained with reinforcement learning using self-feedback: the model writes a first version, checks its own output, finds what's broken and iterates — aimed at end-to-end project generation rather than single-turn coding. Chinese coverage by 量子位 adds details I couldn't verify against official pages: the update is tuned for DeepSeek Harness 0.2.1-rc.1 and above (with special handling for Tencent CodeBuddy and WorkBuddy), a 6-yuan credit for first-time Harness logins, and community demos suggesting noticeably stronger 3D generation. Weights are on Hugging Face under the same open terms as the series.

My read after a day of third-party reports, without running it myself: the output-brevity claim is the number to watch, because verbosity is what quietly inflates agent bills — we saw the same effect quantified in Ling-3.1-flash's pricing math, where a chattier model at a cheaper token rate still lost on total task cost. If DeepSeek genuinely cut tokens and improved scores, the 1017 snapshot is a strict upgrade; treat early community hype on visual quality as unverified until more users publish consistent results.

Who should actually act on this

Wire it up before Nov 1 if you're building search-backed agents — the free window lets you measure real hit rates and latency on your workload without spending anything, and the per-success-request pricing means you can let the agent retry aggressively. Stay on token-only billing if your agent mostly synthesizes from context it already has; paying per search only pays off when grounded freshness actually changes answers. For self-hosted deployments that avoid cloud APIs entirely, our DeepSeek 4.1 Flash deployment guide covers that path — note that self-hosted weights won't include DeepSeek's proprietary search stack, so the Search API is a cloud-side feature.

Caveat (Oct 18, 2026): pricing and model behavior are as published in DeepSeek's official API docs on Oct 17–18, 2026; the free-search window ends Nov 1, 2026 at 00:00 Beijing time. Harness tuning details and community quality claims are credited to 量子位's coverage and not independently verified.