DeepSeek Releases V4.1-Flash-1017 Under MIT and Prices Its Search API at $0.75 per 1,000 Calls

DeepSeek has released V4.1-Flash-1017, an update to its Flash model, with open weights under an MIT licence and a new Search and Fetch API for developers. The weights went up on Hugging Face and ModelScope on October 17, and the new checkpoint is also live in DeepSeek's website, app and API (OpenRouter, Data Studios).

What changed

The model card describes it as a minor update to V4.1-Flash. The context window is 1 million tokens, with a maximum output of 384,000 tokens on DeepSeek's pricing page; the card recommends allowing 256,000 tokens or more of output, a recommendation rather than a conflicting limit. On the API the model is called deepseek-flash, and the older names deepseek-v4-flash and deepseek-v4-flash-vision-exp are now served by V4.1-Flash at Flash pricing, so existing apps switch automatically.

Search and Fetch pricing

The headline addition is a Search and Fetch API, powered by DeepSeek's own search engine and priced at $0.75 per 1,000 successful requests for each endpoint, so 10,000 search calls cost $7.50. Calls made server-side through DeepSeek Harness, Claude Code or Codex are free for now; billing begins at 00:00 Beijing time on November 1, 2026. One outlet reported a limit of five requests per second per account, which DeepSeek's pricing page does not confirm.

Token prices

Per million tokens	Off-peak	Peak (weekdays 01:00-04:00 and 06:00-10:00 UTC)
Input, cache hit	$0.003	$0.006
Input, cache miss	$0.15	$0.30
Output	$0.60	$1.20

Peak hours exclude Chinese public holidays. For developers in India, the peak windows fall between 6:30 and 9:30 a.m. and 11:30 a.m. and 3:30 p.m. IST, so scheduling batch jobs outside those hours halves the bill.

Benchmarks, all from DeepSeek

The model card compares Flash-1017 with the previous Flash on 11 benchmarks, and the new version scores higher on every one. These are DeepSeek's own figures.

Benchmark	Previous Flash	Flash-1017
Humanity's Last Exam (no tools)	36.8	42.6
Terminal-Bench 4.0	31.2	42.3
DeepSWE v1.1	74.2	77.0
ProgramBench	20.3	29.1
NL2Repo	65.4	68.2
BabyVision	89.6	92.5
ZeroBench	49.0	51.0

Four more tests (Automation-Bench, Agents' Last Exam, Apex Bench and Chartography) also improve. The card also compares the model with Opus 5, GPT-5.6 Sol and GLM-5.3; AI Weekly reports Opus 5 at 74.0 and GPT-5.6 Sol at 73.0 on DeepSWE, with Opus 5 leading Humanity's Last Exam at 56.3. Terminal-Bench results depend on the harness, and DeepSeek says it used Claude Code's. Third-party listings such as LLM Stats carry the same figures.

What is disputed

AI Weekly calls the model a 552-billion-parameter mixture-of-experts design. The Hugging Face page lists about 763 billion in stored tensor size and does not state a parameter count. Stored size can differ from the active architecture, so both numbers may be right in different senses, but neither is confirmed by DeepSeek as the model's size.

Why it matters

Open weights under MIT let companies host and modify the model without licence fees, which matters for Indian startups and enterprises that want to keep data on their own servers. The search API adds a paid grounding service that competes with search tools from other model makers, and very cheap cached tokens favour agent workloads that reuse long prompts.

What to watch next

Independent benchmark runs, real bills once search billing starts on November 1, and whether DeepSeek clarifies the model size.

Related on Gadgets365

• Meta releases Muse Wonder and prices its API at $4 per million input tokens
• OpenAI starts hiding watermarks in ChatGPT and Codex text for EU users
• Meta says Muse Spark helped settle five open maths problems, but a critic disputes half