TL;DR. Since 17 October, deepseek-flash answers with DeepSeek-V4.1-Flash-1017, not the 10 September build. The price hasn't moved ($0.15 in, $0.60 out off-peak). There's no dated model id to stay on the old build through the API. The MIT weights are on Hugging Face, but this repo ships without the Jinja chat template the older one gained on 1 October. Search and Fetch cost $0.75 per 1,000 calls, with billing on server-side tools starting 31 October at 16:00 UTC.
If your agent called deepseek-flash yesterday and again this morning, it didn't talk to the same model. DeepSeek's pricing page now lists the official version behind that name as DeepSeek-V4.1-Flash-1017, and the change log entry for 17 October calls it an upgrade. Nothing in your request had to change. That's the problem, if you run evals.
It isn't the first time DeepSeek has swapped the build behind a stable name (we wrote up the 0731 swap in summer). This one looks like a good trade on paper. It still deserves a rerun before you trust it.
What moved behind the name, and what didn't
The price didn't. The pricing page still shows $0.15 per million input tokens on a cache miss and $0.60 out, off-peak, doubled during the weekday peak windows (01:00 to 04:00 and 06:00 to 10:00 UTC). Context stays at 1M with 384K out. So this isn't a repricing; it's a different model at the old rate.
The scores are where DeepSeek puts the weight. The model card's own table, 1017 against the September build: Terminal-Bench 4.0 goes from 31.2 to 42.3, ProgramBench from 20.3 to 29.1, DeepSWE v1.1 from 74.2 to 77.0, and HLE without tools from 36.8 to 42.6 (44.0 on the text-only subset). Those are vendor numbers on the vendor's harness. We haven't rerun any of them, and an eleven point jump on a terminal benchmark in a point release is the kind of thing we'd want to see reproduced before it goes into a slide.
The claim we'd test first is the boring one. DeepSeek says 1017 finishes tasks faster with shorter outputs. If that's true, your bill drops at the same per-token price, because output is the expensive side at four times input. But the announcement gives no figure for how much shorter. Log output tokens per task for a week before and after today and you'll have a better number than anything in the post.
Why you can't pin the September build
Some vendors give you dated snapshot ids you can hold onto. DeepSeek's API doesn't, at least not today. The pricing page lists two model names, deepseek-flash and deepseek-v4-pro, and the old deepseek-v4-flash names are still routed to Flash at Flash prices. A dated deepseek-v4.1-flash-1017 id showed up briefly in DeepSeek's model list overnight, according to an automated tracker account on X, and the API now rejects that name. So if your regression suite assumed deepseek-flash meant one fixed thing, that assumption broke this morning, silently.
The only way to keep the 10 September build is to host it yourself. Both repos sit on Hugging Face under MIT, and each weighs 510.3 GB (we read the sizes from the Hugging Face API on 17 October). That's a multi-GPU server, not a fallback you spin up on a laptop. For most teams, the realistic move is to accept 1017, rerun what matters and record the build name next to every eval result from now on.
One snag for people serving the weights. The September repo gained a chat_template.jinja on 1 October, and the 1017 repo doesn't have one. Its card says so plainly and points to a Python reference encoder in the repo, plus a separate toolkit called deepseek-recipe, for tool calls and thinking mode. If your vLLM or SGLang setup leans on the tokenizer's built-in template, check what it's doing before you point traffic at 1017. We don't know whether the September template is safe to reuse with the new build, and DeepSeek doesn't say.
Search and Fetch, and the 31 October date
The other half of today's entry is DeepSeek's own search engine, opened to developers. Search returns structured web results and Fetch returns one page as Markdown. Both cost $0.75 per 1,000 successful calls on the pricing page, which also says calls made through server-side tools in DeepSeek's harness and in third-party coding agents are free until 00:00 Beijing time on 1 November (16:00 UTC on 31 October). Launch coverage cites a cap of 5 queries per second per account, a figure we couldn't find on DeepSeek's own pages.
Our take: if an agent you run quietly started using DeepSeek's server-side search this week, it's free for two more weeks and then it's metered. Set an alert on that line item now. Whether the index is any good is a separate question we can't answer from a change log, and "hundreds of billions of pages" isn't a quality metric.
We've added 1017 to our open-weight models tracker. The price row in our LLM API pricing tracker doesn't change, since the rates are the same.
FAQ
Did DeepSeek change the price of deepseek-flash?
No. It's still $0.15 per million input tokens on a cache miss, $0.003 on a cache hit and $0.60 per million output tokens off-peak, with double rates in the weekday peak windows. Only the model behind the name changed.
Can I keep using the 10 September build of V4.1-Flash?
Not through the API, as of 17 October. There's no dated id for it. You can self-host the original deepseek-ai/DeepSeek-V4.1-Flash repo, which is 510.3 GB and MIT licensed.
What does the DeepSeek Search API cost?
$0.75 per 1,000 successful requests, the same for Fetch. Calls through server-side tools are free until 31 October 2026 at 16:00 UTC, when billing starts.
Does V4.1-Flash-1017 work with the Hugging Face chat template?
The 1017 repo doesn't include a Jinja chat template. DeepSeek points to a Python reference encoder in the repo and to its deepseek-recipe toolkit for formatting tool calls and thinking mode.
Sources: DeepSeek, API change log (17 October 2026), V4.1-Flash-1017 announcement and models and pricing; Hugging Face model cards for V4.1-Flash-1017 and V4.1-Flash, with sizes and commit history read from the Hugging Face API on 17 October 2026; independent: AI Weekly and KuCoin News.
