DeepSeek V4.1 Flash 1017, released October 17, 2026, is an open-weights update that scores 77.0 on DeepSWE v1.1, up from 74.2, while cutting DeepSeek’s reported average cost per coding task from about $0.50 to just under $0.20. Token prices stay at $0.30 input and $1.20 output per million at peak, and DeepSeek opened its Search and Fetch APIs at $0.75 per 1,000 requests.
Key takeaways
- DeepSWE v1.1 score rose to 77.0, ahead of Claude Opus 5’s 74.0.
- Average cost per DeepSWE task fell from about $0.50 to under $0.20.
- Search and Fetch APIs cost $0.75 per 1,000 successful requests from November 1.
What is DeepSeek V4.1 Flash 1017?
DeepSeek V4.1 Flash 1017 is a new checkpoint of the V4.1 Flash model DeepSeek shipped in September. It replaces the old model on the website, the app and the API under the same deepseek-flash name. The weights are on Hugging Face under an MIT license, and the price did not change.
DeepSeek’s own changelog calls it “a new performance–cost Pareto optimum across multiple harnesses.” The model card is more modest. It describes the release as “a minor update to DeepSeek-V4.1-Flash.”
The pitch is efficiency, not size. DeepSeek says the update shortens task runtime and output length, so the same job burns fewer tokens. The Hugging Face files list 763 billion parameters and a recommended context window of 1 million tokens.
We covered the original launch in DeepSeek V4.1 Flash: 552B Open Weights, 75% Less KV Cache. Five weeks later, the new fact is the cost per finished task, not the architecture.
What changed under the hood?
According to The Decoder, DeepSeek trained the update with reinforcement learning from self-feedback, where the model’s own confidence scores or critiques act as the reward signal. It was also tuned for DeepSeek Harness and for Tencent’s CodeBuddy and WorkBuddy tools.
DeepSeek recommends Harness version 0.2.1-rc.1 or later. The legacy API names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work, but the pricing page says those models are retired and now route to V4.1 Flash 1017.
How does DeepSeek V4.1 Flash 1017 score against rivals?
On DeepSeek’s own numbers, V4.1 Flash 1017 beats Claude Opus 5, GPT-5.6 Sol, Kimi K3 and GLM-5.3 on four agentic benchmarks, including DeepSWE and Automation-Bench. It still trails Opus 5 on Terminal-Bench 4.0, ProgramBench, NL2Repo and Humanity’s Last Exam. None of these scores has been independently verified yet.
| Benchmark | V4.1 Flash 1017 | V4.1 Flash (Sept) | Kimi K3 | GLM-5.3 | GPT-5.6 Sol | Claude Opus 5 |
|---|---|---|---|---|---|---|
| DeepSWE v1.1 | 77.0 | 74.2 | 67.5 | 66.9 | 73.0 | 74.0 |
| Terminal-Bench 4.0 | 42.3 | 31.2 | 12.6 | 37.9 | 39.9 | 51.8 |
| Automation-Bench v1.0.6 | 60.5 | 54.8 | 46.7 | 48.8 | 45.8 | 50.3 |
| Agents’ Last Exam | 35.4 | 31.8 | 27.6 | 28.5 | 26.7 | 28.6 |
| Apex Bench v1.1 | 72.1 | 64.6 | 47.8 | 56.6 | 54.3 | 65.8 |
| ProgramBench | 29.1 | 20.3 | 17.5 | 19.0 | 23.0 | 37.0 |
| NL2Repo-Bench | 68.2 | 65.4 | 58.0 | 58.0 | 56.8 | 75.3 |
| HLE (no tools) | 42.6 | 36.8 | 43.5 | 42.0 | 44.5 | 56.3 |
The biggest jump is on Terminal-Bench 4.0, up 11.1 points to 42.3. ProgramBench rose 8.8 points. Those are large moves for something DeepSeek itself labels a minor update.
Here is the skeptical read. The comparison set leans on older frontier models. Claude Opus 5 is no longer Anthropic’s top model, and the table leaves out GPT-6 and Gemini 4 entirely. Beating last season’s leaders on a vendor-picked benchmark list is a weaker claim than it looks.
How much does DeepSeek V4.1 Flash 1017 cost?
Nothing changed on the price sheet. DeepSeek’s pricing page lists deepseek-flash at $0.30 per million cache-miss input tokens and $1.20 per million output tokens at peak. Off-peak rates are half that: $0.15 input and $0.60 output. Cached input costs $0.006 per million at peak.
| Charge (per 1M tokens) | deepseek-flash peak | deepseek-flash off-peak | deepseek-v4-pro peak | deepseek-v4-pro off-peak |
|---|---|---|---|---|
| Input, cache hit | $0.006 | $0.003 | $0.044 | $0.022 |
| Input, cache miss | $0.30 | $0.15 | $1.32 | $0.66 |
| Output | $1.20 | $0.60 | $3.96 | $1.98 |
The detail US buyers should notice is the clock. DeepSeek defines peak hours as 01:00–04:00 and 06:00–10:00 UTC on weekdays. That is overnight and early morning on the US East Coast. A New York team working 9 a.m. to 6 p.m. pays the off-peak rate almost all day.
Why the cost per task matters more than the token price
Agents are billed by the token but judged by the finished job. A model that thinks less and writes shorter answers gets cheaper without a price cut. According to The Decoder’s reading of DeepSeek’s chart, the 1017 checkpoint at its top setting averages just under $0.20 per DeepSWE task.
- V4.1 Flash 1017, top setting: 77.0% at just under $0.20 per task.
- V4.1 Flash 1017, low setting: above 72% at about $0.10 per task.
- September V4.1 Flash: 74.2% at about $0.50 per task.
- Western rivals in DeepSeek’s chart: roughly $2 to more than $10 per task.
Those figures come from DeepSeek, not from an independent lab. Still, if they hold, a team running 10,000 agent tasks a month would spend about $2,000 instead of $5,000 on the same model.
What are the DeepSeek Search and Fetch APIs?
The Search and Fetch APIs open DeepSeek’s in-house search engine to any developer. Search returns ranked web results for a query. Fetch returns a single page as Markdown. Each costs $0.75 per 1,000 successful requests, and failed or empty calls are free, according to DeepSeek’s documentation.
DeepSeek’s changelog says the engine “indexes hundreds of billions of web pages.” It is the same index that powers search inside DeepSeek’s own app and API.
- Results: 1 to 50 per query, with a 0–3 rerank score and a 0–4 authority tier.
- Operators:
site:,lang:,after:,before:and exact-phrase matching. - Rate limit: 5 queries per second per account for each API.
- Free window: server-side calls through DeepSeek Harness, Claude Code or Codex are free until billing starts at 00:00 Beijing time on November 1, 2026.
The 5-per-second cap is the catch. That is fine for a coding agent and far too low for a consumer search product. DeepSeek is selling a tool for agents, not a full replacement for a web search API.
Who wins and who loses from DeepSeek V4.1 Flash 1017?
Developers running high-volume coding agents win first, because their cost per finished task drops without a migration. Tencent gains from a model tuned for its CodeBuddy tools. US labs selling agentic coding at several dollars a task face more pricing pressure, even if their top models still score higher.
The winners
Startups that resell agents keep more margin. An MIT-licensed checkpoint also lets them self-host, which caps their exposure to DeepSeek’s API. That matters after DeepSeek raised some API prices by up to 1,100% in August, as we reported in DeepSeek Price Increase: Up to 1,100% Overnight.
DeepSeek itself wins distribution. Search calls routed through Claude Code and Codex put its engine inside rival tools. That is a cheap way to build usage ahead of the funding round we covered in DeepSeek Funding Round Tops $12B at a $74B Valuation.
The losers
Closed US labs lose pricing power at the low end of agent work. When a free-to-host model lands within a few points of Opus 5 on coding, buyers ask why routine tasks cost 10 times more. Other open-weight players feel it too. Meta’s Muse Spark open weights arrived one day earlier and now face a cheaper benchmark rival.
There is also a compliance cost. US companies in regulated sectors may not be allowed to send code or search queries to a Chinese API. For them, the only path is self-hosting, and Hugging Face showed no inference provider serving the model at launch.
What to watch next
The next two weeks decide whether this is a real cost shift or a benchmark headline. Three signals matter most: the Search API billing start, independent benchmark results, and whether US inference hosts pick up the weights. Each one tests a different part of DeepSeek’s claim.
- November 1, 2026 (00:00 Beijing time): free server-side Search and Fetch calls end and per-request billing starts.
- Independent benchmarks: The Decoder noted no third-party results at launch. Watch for Terminal-Bench and SWE-style leaderboards to confirm or cut the gains.
- Inference providers: a US host listing the model on Hugging Face would open it to buyers who cannot use DeepSeek’s API.
FAQ
Is DeepSeek V4.1 Flash 1017 open source?
The weights are open under an MIT license on Hugging Face and ModelScope. You can download, modify and run them commercially.
How much does DeepSeek V4.1 Flash 1017 cost on the API?
At peak, $0.30 per million cache-miss input tokens and $1.20 per million output tokens. Off-peak prices are half.
Do I need to change my code to use it?
No. DeepSeek says the calling method and pricing are unchanged. The deepseek-flash model name now points to the 1017 checkpoint.
Is it better than Claude Opus 5?
On DeepSeek’s numbers, it beats Opus 5 on DeepSWE, 77.0 to 74.0, but trails on Terminal-Bench 4.0, 42.3 to 51.8. The scores are not independently verified.
What does the DeepSeek Search API cost?
$0.75 per 1,000 successful requests for Search and the same for Fetch. Calls through DeepSeek Harness, Claude Code or Codex are free until November 1, 2026.
When are DeepSeek’s off-peak hours?
Any time outside 01:00–04:00 and 06:00–10:00 UTC on weekdays, plus weekends and Chinese public holidays. That covers most of the US workday.
The bottom line
DeepSeek V4.1 Flash 1017 does not move the frontier. It moves the price of getting work done. A self-reported 60% drop in cost per coding task, at unchanged token prices, is the kind of gain that compounds for anyone running agents at scale.
Our verdict: worth testing now for routine coding and automation work, especially during off-peak US hours. Wait for independent benchmarks before trusting it on hard terminal tasks, where Opus 5 still leads by 9.5 points. For investors, the signal is plain. The cheapest tier of agent work keeps getting cheaper, and closed labs will have to win on the hard tasks.
Free guide · 12 pages
AI Tools Every Investor Should Use in 2026
10 tools for research, stock signals, charts and crypto, what each one costs, three ready-made stacks from $0 a month and 7 copy-paste prompts for filings and earnings calls. Enter your email and we will send it to you.
Plus one email a week: AI Money This Week, every Sunday. Unsubscribe anytime.
Sources
- DeepSeek-V4.1-Flash-1017 model card, Hugging Face
- DeepSeek API change log, October 17, 2026
- DeepSeek API models and pricing
- DeepSeek Search and Fetch API guide
- The Decoder: DeepSeek V4.1 Flash 1017 beats larger rivals in agentic tasks
Wealth Engine researches and drafts with AI tools and checks every figure against the sources above. How we report.
Leave a Reply