DeepSeek ships V4.1 Flash 1017 and opens its search engine to the API

What happened

DeepSeek's API change log for October 17, 2026 has two entries. The first upgrades DeepSeek V4.1 Flash, released on September 10, to a minor update called DeepSeek-V4.1-Flash-1017. The web chat and the apps switched to it the same day, and the API kept its model name and its prices. The second entry opens DeepSeek's own search engine, which already powers search in its chat, to developers as a Search API and a Fetch API. The weights of the new model went up on Hugging Face under the MIT license, and by October 20 the model page showed close to 18,000 downloads.

The model: same name, same price, better agent scores

No code changes are needed. The model name stays deepseek-flash, and the older names deepseek-v4-flash and deepseek-v4-flash-vision-exp are now routed to 1017 as well, billed at the Flash price. DeepSeek says the update improves reasoning, coding, agent tasks and image understanding, and also cuts how long a task runs and how much text the model writes, so the same job costs less.

DeepSeek's own numbers against the September release:

These are the vendor's tests, run in harnesses it chose (Claude Code for Terminal-Bench 4.0, Mini-SWE-V2 for DeepSWE). They show a large jump on terminal and programming tasks over September, and a model that still trails Opus 5 on Terminal-Bench, ProgramBench and HLE while beating it on DeepSWE and on DeepSeek's automation and Apex tests.

The rest of the spec comes from the model card and the pricing page. It is a mixture-of-experts model with 552B backbone parameters, of which 8B are active per token while it reads the prompt and 16B while it writes. It takes text and images, holds a context of up to 1M tokens and can return up to 384K tokens. DeepSeek recommends a temperature of 1.0 and says the update was tuned for its own DeepSeek Harness and for Tencent's CodeBuddy and WorkBuddy.

Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, except Chinese public holidays. Everything else, weekends included, is billed at the off-peak rate.

The new Search and Fetch APIs

• Search (POST /v1/tools/search) takes a query of up to 128 characters and returns 1 to 50 results. Each comes with a short snippet for display, a 600 to 800 character summary meant for a model to read, the full page text if you ask for it, a date, and a relevance score from 0 to 3, where 1.5 and above counts as a good result.
• Operators work in the query: site:, -site:, exact phrases in quotes, lang:, after: and before:, so an agent can ask for "pages on github.com in English published after 2025" in one call.
• Fetch (POST /v1/tools/fetch) takes one URL and returns the page as Markdown, ready to put into a prompt.
• Price: $0.75 per 1,000 successful requests for each, billed separately from tokens. A failed request costs nothing.
• Free until November 1 in agent tools. When DeepSeek models run inside DeepSeek Harness, Claude Code or Codex, the tools' built-in web search already goes through these APIs, and those calls are free until 00:00 Beijing time on November 1, 2026 (16:00 UTC on October 31).

DeepSeek says its engine indexes hundreds of billions of pages and filters SEO spam and low-quality AI-written pages. That is its claim, and the scores in the results are the easiest way to check it on your own queries.

What it means for you

1. Re-run your evals. An app that calls deepseek-flash switched models on October 17 without a code change. Check a sample of real tasks, especially anything that parses the output, since answers are now shorter.
2. Schedule heavy jobs off-peak. Batch work moved outside the peak windows costs half.
3. Price the search before November 1. Running DeepSeek inside Claude Code or Codex? Count how many web searches a typical session makes this week while they're free. At $0.75 per 1,000, a session that makes 40 searches adds three cents.
4. Check where your data goes. DeepSeek's privacy policy says personal data is processed and stored in China, which many companies' rules don't allow. The MIT-licensed weights get around that for teams that can host a model this size themselves.

Our DeepSeek review keeps the full price table and the privacy details. For prompts that make a coding agent work in small, checkable steps, see the Claude Code prompts, which work the same way with a DeepSeek model behind them.

Try it

I'm comparing two answers to the same task from two versions of a model. For each pair I paste, tell me:
1) which answer is correct, or whether both are,
2) what one includes that the other leaves out,
3) whether the shorter one dropped anything I actually need,
4) which one I'd ship, in one line.
Be strict about facts and code that would not run. Don't reward length.

Task: [the task or prompt]
Answer A: [old output]
Answer B: [new output]

Before you answer, search the web. Write your search queries with operators instead of plain questions:
- use site: to go to the source (docs, GitHub, the vendor's site) when the question is about a product,
- use after: with a date from the last 12 months when the answer can go out of date,
- use quotes for exact error messages or names.
Run at most 5 searches. Then answer with a list of the sources you used, each with its date, and say which claim each one supports.

My question: [your question]

Sources

• DeepSeek API Docs, Change Log, entries of October 17 and September 10, 2026
• DeepSeek API Docs, Models & Pricing and Search & Fetch guide, checked October 20, 2026
• Hugging Face, DeepSeek-V4.1-Flash-1017 model card and DeepSeek-V4.1-Flash model card, checked October 20, 2026