DeepSeek released V4.1-Flash-1017 on Oct 17, 2026, a post-training refresh of its low-cost Flash model, live now under deepseek-flash name with API prices unchanged.
What happened
DeepSeek released DeepSeek-V4.1-Flash-1017 on October 17, 2026, an updated version of the low-cost Flash model it launched on September 10. The new checkpoint is already live on DeepSeek’s website, app, and API under the existing deepseek-flash model name, with API prices unchanged. DeepSeek also published the model weights under the MIT license on Hugging Face and ModelScope.
The update is a post-training refresh, not a new architecture. If you already call deepseek-flash, you are already on the new snapshot — no code change is needed. Older model names still work, but the models behind them have been retired and requests are billed at the Flash price.
On DeepSeek’s own benchmark table, all 11 reported scores improved. The headline numbers: DeepSWE v1.1 rose from 74.2 to 77.0, Terminal-Bench 4.0 from 31.2 to 42.3, and ProgramBench from 20.3 to 29.1. In DeepSeek’s comparison, the new model beats GPT-5.6 Sol on nine of eleven tests but trails Claude Opus 5 on seven, including Humanity’s Last Exam (42.6 versus 56.3). Every one of these scores is DeepSeek’s own, produced on test harnesses it chose, and none had been independently reproduced at the time of writing.
DeepSeek also says the update finishes tasks faster and produces shorter outputs, which would lower a bill even at an unchanged price per token. The company did not publish task-time or output-length figures, so that saving cannot be sized yet.
Beyond the model, DeepSeek opened its Search and Fetch APIs to all developers. Each endpoint costs $0.75 per 1,000 calls, with calls made through server-side tools in DeepSeek Harness, Claude Code, and Codex free until per-request billing starts on November 1, 2026, Beijing time. DeepSeek says the APIs are powered by an in-house search engine serving hundreds of millions of requests a day. The model is tuned for DeepSeek Harness and for Tencent’s CodeBuddy and WorkBuddy.
Why it matters for small business
For a small business, the story here is agent economics. A 1-million-token context window at $0.15 per million input tokens (off-peak) makes long, multi-step agent runs affordable: triaging a support inbox, working through a pile of invoices, or letting a coding agent grind through a backlog of small fixes. Shorter outputs and faster completion would push the effective cost lower still, if DeepSeek’s claims hold up in real use.
The silent upgrade cuts both ways. The same API name means zero migration work, but it also means results from September tests no longer describe the model you are calling today. Re-run your own evaluations on your own tasks — your invoice formats, your support tickets, your codebase — before trusting the new snapshot with anything customer-facing.
The Search and Fetch APIs matter because many small-business agents need fresh information from the web: competitor prices, supplier pages, current documentation. Getting the model and a web-search layer behind one API key removes a piece of plumbing that small teams usually have to build themselves.
The open weights are genuine option value, but read the fine print: 510 GB of weights across 48 files. Self-hosting is a project for a team with serious GPU capacity, not a weekend task. For most small businesses, the API is the practical route, and the MIT license mostly matters as a guarantee that the model cannot be quietly taken away.
Stay cautious. Vendor benchmarks are marketing until independent labs reproduce them, DeepSeek’s own report notes robustness limits that are not yet fully characterized, and recent coverage has flagged operational hiccups. Keep a fallback model configured, and never let an agent take an irreversible action — sending money, emailing customers, deleting data — without a human check.
Pricing, availability, and what to do next
Availability: Live now on DeepSeek’s website, app, and API as deepseek-flash. Weights are available on Hugging Face and ModelScope under the MIT license.
Pricing (DeepSeek’s published rates): $0.15 per million input tokens off-peak ($0.30 at peak), $0.003 per million cached input tokens, and $0.60 per million output tokens off-peak. Peak rates are double, during peak hours of 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday, excluding Chinese public holidays.
What to do next:
- Re-run your own tests on
deepseek-flashbefore relying on September results. - Try the Search and Fetch APIs before November 1, while server-side calls are still free, and see whether they replace a separate search API in your agents.
- Set spend caps and per-task token budgets; cheap per-token pricing still adds up on long agent loops.
- Keep a second model as a fallback for outages and for tasks where Flash’s quality falls short.
Sources
- TECHi, “DeepSeek’s Flash update claims gains on 11 benchmarks at the same price”: https://www.techi.com/deepseek-v4-1-flash-1017-update-benchmarks-price/
- Dealroom, “DeepSeek moves up the stack: V4.1-Flash update ships alongside its own Search and Fetch APIs for agents”: https://dealroom.co/news/163453-deepseek-moves-up-the-stack-v4-1-flash-update-ships-alongside-its-own-se/
