DeepSeek V4.1 Flash 1017 went live on October 17, 2026, and DeepSeek didn’t touch the price. The upgraded model now answers every API call to deepseek-flash, plus DeepSeek’s web chat and app, at the same rates as before: $0.15 per million input tokens and $0.60 per million output tokens off-peak, or $0.30 and $1.20 at peak. The weights are on Hugging Face under the MIT license, the same day.
DeepSeek’s changelog calls it a “performance–cost Pareto improvement”: better scores, shorter outputs and faster task runs than the September 10 version of V4.1 Flash. Those are DeepSeek’s own numbers. Two details matter more for anyone outside China. DeepSeek’s peak hours mostly fall overnight in the US, so a normal American workday is billed at the cheaper rate. And the rival models in DeepSeek’s benchmark table aren’t the newest ones from OpenAI or Anthropic.
DeepSeek V4.1 Flash 1017 price: what you actually pay
Here are DeepSeek’s Models & Pricing page rates, checked the morning of October 17, 2026. Off-peak is half the peak price.
| Per 1M tokens (DeepSeek API, Oct 17) | Off-peak | Peak |
|---|---|---|
| Input, cache hit | $0.003 | $0.006 |
| Input, cache miss | $0.15 | $0.30 |
| Output | $0.60 | $1.20 |
Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday, excluding Chinese public holidays. Weekends are off-peak all day. On US Eastern daylight time, that works out to 9 PM to midnight Sunday through Thursday, plus 2 AM to 6 AM on weekdays. After clocks change on November 1, it shifts an hour earlier (8 to 11 PM and 1 to 5 AM ET). So a developer in New York or San Francisco who runs jobs during business hours pays the off-peak rate.
The model also covers the specs people ask about. It has a 1-million-token context window and a maximum output of 384K tokens, and thinking mode is on by default (non-thinking is available). It handles vision input, JSON output and tool calls, and supports both OpenAI-format and Anthropic-format endpoints, including the Responses API. The concurrency limit is 2,500. The old model names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work, but DeepSeek says those models are retired and their requests are now served by V4.1-Flash-1017 at the Flash price.
For comparison, DeepSeek’s bigger V4-Pro stays at $0.66 input and $1.98 output off-peak ($1.32 and $3.96 at peak), and it has no vision support.
The benchmark table, and who’s missing from it
The Hugging Face model card puts the update next to the September version and four rivals. Every number below comes from DeepSeek’s table. DeepSeek doesn’t say who ran the rival scores, so these aren’t independent results.
| Benchmark (DeepSeek’s table) | V4.1-Flash-1017 | V4.1-Flash (Sep) | GPT-5.6 Sol | Opus 5 |
|---|---|---|---|---|
| HLE, no tools | 42.6 | 36.8 | 44.5 | 56.3 |
| Terminal-Bench 4.0 | 42.3 | 31.2 | 39.9 | 51.8 |
| DeepSWE v1.1 | 77.0 | 74.2 | 73.0 | 74.0 |
| ProgramBench | 29.1 | 20.3 | 23.0 | 37.0 |
| NL2Repo-Bench | 68.2 | 65.4 | 56.8 | 75.3 |
| Automation-Bench v1.0.6 | 60.5 | 54.8 | 45.8 | 50.3 |
| Agents’ Last Exam | 35.4 | 31.8 | 26.7 | 28.6 |
| Apex Bench v1.1 | 72.1 | 64.6 | 54.3 | 65.8 |
DeepSeek ran Terminal-Bench 4.0 in the Claude Code harness and DeepSWE in Mini-SWE-V2. The jump over the September model is real on DeepSeek’s own runs: Terminal-Bench 4.0 climbs 11.1 points and ProgramBench climbs 8.8. Against Anthropic’s Claude Opus 5, the new Flash wins on DeepSWE, Automation-Bench, Agents’ Last Exam and Apex Bench. It loses on HLE, Terminal-Bench 4.0, ProgramBench and NL2Repo, and Opus 5 also scores slightly higher on DeepSeek’s three tool-assisted vision tests (Chartography, BabyVision and ZeroBench-main). The table also lists Moonshot’s Kimi K3 and Z.ai’s GLM-5.3. The new Flash beats each of them on every row where they’re scored, except one: Kimi K3 edges it on HLE without tools, 43.5 to 42.6.
Here’s the catch: Opus 5 and GPT-5.6 Sol aren’t the current flagships. Anthropic released Claude Opus 5.5 on September 22, and OpenAI’s lineup has moved on to the GPT-6 family. Neither appears in DeepSeek’s table. So “beats GPT and Claude” claims about this release only hold against older models.
Open weights, architecture and the new Search API
The weights are on Hugging Face as deepseek-ai/DeepSeek-V4.1-Flash-1017 and on ModelScope, under MIT, which allows commercial use. DeepSeek calls 1017 a “minor update” to V4.1 Flash. According to that model’s card, V4.1 Flash is a multimodal mixture-of-experts model with 552 billion backbone parameters. It activates 8 billion per token during prefill and 16 billion during decode, and uses 384 routed experts with 6 active per token, plus a 196-billion-parameter “Engram” memory. DeepSeek doesn’t publish a minimum hardware spec for running it yourself. Its reference inference example uses 8-way tensor parallelism, and the release ships without a standard chat template; DeepSeek provides its own encoding code instead.
The same release opens DeepSeek’s in-house Search and Fetch APIs to everyone, at $0.75 per 1,000 successful requests for each. When you use the DeepSeek API through DeepSeek Harness, Claude Code or Codex, the server-side search calls are free until billing starts at midnight Beijing time on November 1 (noon ET on October 31).
Next to watch: independent leaderboard results for 1017, and whether DeepSeek adds Opus 5.5 or GPT-6 to its comparison. We cover the open-weight race in our Muse Spark Open and Mistral Large 4 reports.
Frequently asked questions
How much does DeepSeek V4.1 Flash 1017 cost?
On DeepSeek’s API, checked October 17, 2026: $0.15 per million input tokens (cache miss) and $0.60 per million output tokens off-peak, or $0.30 and $1.20 at peak. Cache hits cost $0.003 off-peak and $0.006 at peak. DeepSeek says the price didn’t change with this update.
When are DeepSeek’s peak hours in US time?
Peak is 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays. On Eastern daylight time that’s 9 PM to midnight Sunday through Thursday and 2 AM to 6 AM on weekdays. After November 1 it’s one hour earlier. Everything else, including weekends, is off-peak.
Is DeepSeek V4.1 Flash 1017 open source?
The weights are on Hugging Face (deepseek-ai/DeepSeek-V4.1-Flash-1017) and ModelScope under the MIT License, which allows commercial use.
What is the DeepSeek V4.1 Flash 1017 context window?
1 million tokens, with a maximum output of 384K tokens on the API. It accepts image input as well as text.
How do I use DeepSeek V4.1 Flash 1017 in the API?
Use the model name deepseek-flash. The older names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted and are now served by V4.1-Flash-1017 at the Flash price.
Does DeepSeek V4.1 Flash 1017 beat Claude and GPT?
On DeepSeek’s own table it beats Claude Opus 5 on DeepSWE, Automation-Bench, Agents’ Last Exam and Apex Bench, and trails it on HLE, Terminal-Bench 4.0, ProgramBench, NL2Repo and three vision tests. The table compares it with Opus 5 and GPT-5.6 Sol, not the newer Claude Opus 5.5 or OpenAI’s GPT-6 models.
Sources
All accessed October 17, 2026, between about 8:45 and 9:15 AM ET.
- DeepSeek API Docs, news post “DeepSeek-V4.1-Flash-1017: Wish-Driven Programming Made Real” (Oct 17, 2026; release, open-source links, Search and Fetch pricing): api-docs.deepseek.com/news/news261017
- DeepSeek API Docs, Change Log entry dated 2026-10-17 (pricing unchanged, benchmark list, harness notes): api-docs.deepseek.com/updates
- DeepSeek API Docs, Models & Pricing (prices, peak hours, context, max output, features, concurrency, legacy model names; checked Oct 17, 2026): api-docs.deepseek.com/quick_start/pricing
- Hugging Face model card, deepseek-ai/DeepSeek-V4.1-Flash-1017 (benchmark table vs Kimi K3, GLM-5.3, GPT-5.6 Sol and Opus 5; MIT license; inference notes): huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017
- Hugging Face model card, deepseek-ai/DeepSeek-V4.1-Flash (architecture: 552B backbone, 8B/16B active, 384 routed experts, Engram 196B): huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash
- DeepSeek on X, launch thread (Oct 17, 2026, 4:35 AM ET): x.com/deepseek_ai/status/2111375640717033777
- Anthropic newsroom, “Introducing Claude Opus 5.5” (Sep 22, 2026): anthropic.com/news
Related: Muse Spark Open download, price and hardware · Mistral Large 4 price and open weights · More LLM news
Last updated: October 17, 2026. To report an error, see our corrections page. Articles are drafted with AI assistance and reviewed and edited by an editor; see our editorial policy.
