DeepSeek V4.1-Flash-1017 and the new Search API: what changed, peak hours in IST, and what it costs
DeepSeek swapped the model behind deepseek-flash, opened MIT weights and launched a $0.75-per-1,000 search API. Prices, limits and IST peak hours.
DeepSeek shipped two things on 17 October. The first is DeepSeek-V4.1-Flash-1017, an update to its low-cost Flash model that now sits behind the deepseek-flash API name, with MIT-licensed weights on Hugging Face. The second is a paid web Search and Fetch API built on DeepSeek’s own search engine. If you run coding agents or retrieval pipelines on a tight budget, both are worth a look. A few details matter before you switch: what happens to your existing model names, when peak pricing hits Indian working hours, and how far to trust the benchmark table.
What changed in the API
According to DeepSeek’s change log entry for 17 October, V4.1 Flash has been upgraded to the 1017 version on the API, the web app and the mobile app. The calling method and the price are unchanged. There is no separate dated model ID to call: you reach the new model through the deepseek-flash alias.
DeepSeek’s API docs for the release add two points that affect existing integrations:
deepseek-flashis now served by V4.1-Flash-1017.- The older names
deepseek-v4-flashanddeepseek-v4-flash-vision-expare still accepted, but the models behind them are retired. Requests to those names are answered by V4.1-Flash-1017 and billed at the Flash price.
In practice, if you never pinned a model, your traffic moved to the new checkpoint on Saturday. That is convenient, but it also means there is no way to stay on the previous Flash through DeepSeek’s API. If you have evaluation sets or prompts tuned to the older behaviour, re-run them this week. The model card says the update reduces task execution time and output length, so expect shorter answers, which may help cost but can break parsers that expect a certain shape.
DeepSeek also says its agent tool, DeepSeek Harness, should be on 0.2.1-rc.1 or later for this version.
Prices, and the peak-hour catch for India
DeepSeek’s pricing page lists two models. Prices are US$ per million tokens, shown as off-peak / peak:
| Model | Input, cache hit | Input, cache miss | Output | Concurrency limit |
|---|---|---|---|---|
deepseek-flash (V4.1-Flash-1017) |
0.003 / 0.006 | 0.15 / 0.30 | 0.60 / 1.20 | 2,500 |
deepseek-v4-pro (V4-Pro-0813) |
0.022 / 0.044 | 0.66 / 1.32 | 1.98 / 3.96 | 500 |
The same page gives a 1M-token context window and a 384K-token maximum output, with thinking mode on by default.
The detail that matters most for teams in India is the peak window. DeepSeek charges peak rates from 01:00 to 04:00 and from 06:00 to 10:00 UTC, Monday to Friday, except Chinese public holidays. Converted to IST, that is 06:30–09:30 and 11:30–15:30 on weekdays, which covers a large part of the Indian working day. Peak prices are double the off-peak prices. Weekends and Chinese public holidays are entirely off-peak.
Two practical consequences:
- Batch jobs (overnight evals, document processing, embeddings-style summarisation) cost half as much if you schedule them outside those windows, for example in the Indian evening.
- Interactive products used by Indian customers during office hours will mostly pay the peak rate. Budget on that figure, not the off-peak one.
The new Search and Fetch API
The Search & Fetch guide describes two endpoints on https://api.deepseek.com:
POST /v1/tools/searchtakes a query of up to 128 characters and returns up to 50 results. Each result has a title, URL, a short snippet, a longer model-oriented summary, an optional page content field, a date, and two scores: a relevance score from 0 to 3 and an “authority” score from 0 to 4. Queries supportsite:,-site:, exact phrases, exclusions,lang:andafter:/before:operators.POST /v1/tools/fetchtakes a URL and returns the page as Markdown.
The change log says the engine indexes hundreds of billions of pages and supports multilingual search. The docs do not list which languages, so if you need Hindi, Tamil or other Indian-language results, test with the lang: operator before you rely on it.
Pricing and limits. Each API costs US$0.75 per 1,000 successful requests, whatever the count value. Only HTTP 200 responses with results are billed; empty results, failures and timeouts are free. Both APIs are capped at 5 queries per second per account, separately from your model concurrency, and Fetch also has a per-domain limit. Over the limit you get HTTP 429.
Free until November. Server-side tool calls made through DeepSeek Harness, Claude Code or Codex are currently free. The guide says per-request billing starts at 00:00 Beijing time on 1 November 2026, which is 21:30 IST on 31 October.
For a sense of scale: an agent that runs 20 searches per task would spend US$0.015 per task on search at the listed price. The 5 QPS cap is the more likely constraint for a busy product, so check it against your traffic before you build on it.
The model itself
The model card on Hugging Face calls 1017 “a minor update” to V4.1 Flash. It handles both text and images, and the code and weights are under the MIT licence, which allows commercial use and fine-tuning without a revenue threshold.
The size figures don’t line up across DeepSeek’s pages. DeepSeek’s September announcement of V4.1 Flash describes a 552B-parameter mixture-of-experts model with 8B active parameters for input and 16B for output. The Hugging Face page for the 1017 weights shows 763B parameters in its file metadata. We could not find an explanation on either page, so treat the exact count as unconfirmed. Either way, this is not a model for a single workstation GPU. The card gives no hardware figures. It shows serving examples for vLLM and SGLang, and Hugging Face lists about 19 community-quantised versions for llama.cpp, Ollama and similar tools.
One thing to know if you self-host: the card ships no Jinja chat template. Prompts are built with a Python reference encoder in the repo, and DeepSeek points production users to its deepseek-recipe library.
The benchmarks, read carefully
All of these numbers come from DeepSeek’s own table on the model card. We haven’t seen independent results yet.
| Benchmark | V4.1-Flash-1017 | V4.1-Flash | GPT-5.6 Sol | Claude Opus 5 |
|---|---|---|---|---|
| DeepSWE v1.1 | 77.0 | 74.2 | 73.0 | 74.0 |
| Terminal-Bench 4.0 | 42.3 | 31.2 | 39.9 | 51.8 |
| ProgramBench | 29.1 | 20.3 | 23.0 | 37.0 |
| HLE (no tools) | 42.6 | 36.8 | 44.5 | 56.3 |
| Automation-Bench v1.0.6 | 60.5 | 54.8 | 45.8 | 50.3 |
The card notes that Terminal-Bench was run with the Claude Code harness and DeepSWE with a Mini-SWE-V2 harness. The pattern is clear enough: the update is a solid step over the previous Flash, it leads this table on DeepSWE and Automation-Bench, and Claude Opus 5 stays well ahead on Terminal-Bench, ProgramBench and HLE. The comparison models are not the newest from each vendor, so don’t read this as a ranking of current frontier models.
What to do this week
- If you call
deepseek-flashor the old V4 Flash names: you are already on 1017. Re-run your evals and check for shorter outputs. - If you schedule bulk work: move it outside 06:30–09:30 and 11:30–15:30 IST on weekdays to pay half.
- If you want web search in an agent: try the Search API through Claude Code, Codex or DeepSeek Harness while it is free, and measure result quality for your queries and languages before billing starts on 1 November.
- If you handle customer or regulated data: check where requests are processed and what your contracts and policies allow before sending production data to any third-party API, this one included.
References
- Change Log (2026-10-17: DeepSeek-V4.1-Flash-1017 and Search/Fetch API) — DeepSeek API Docs, accessed 2026-10-18
- DeepSeek-V4.1-Flash-1017 release page (model alias notes) — DeepSeek API Docs, accessed 2026-10-18
- Models & Pricing — DeepSeek API Docs, accessed 2026-10-18
- Search & Fetch API guide — DeepSeek API Docs, accessed 2026-10-18
- deepseek-ai/DeepSeek-V4.1-Flash-1017 (model card) — Hugging Face / DeepSeek, accessed 2026-10-18
- DeepSeek-V4.1-Flash: Smarter, Faster, More Efficient — DeepSeek API Docs, accessed 2026-10-18
This week in AI, every Sunday
One email with the releases that mattered and what they mean in practice. No spam.