DeepSeek V4.1 Flash 1017 Is Out: 77% DeepSWE at $0.20 a Task
DeepSeek has shipped DeepSeek-V4.1-Flash-1017, an update to its Flash model that it says scores higher, finishes agent tasks faster and writes shorter answers, and it has posted the weights under an MIT license. The Chinese AI lab announced the release in a four-part thread on X at 08:35 UTC on Saturday, October 17, 2026 (4:35 p.m. in Beijing). The headline number is cost: on DeepSeek's own DeepSWE chart, the model reaches 77.0% for just under $0.20 per task. DeepSeek also opened its in-house web search engine to outside developers for the first time, through new Search and Fetch APIs priced at $0.75 per 1,000 calls each.
What happened
DeepSeek calls the release "a Pareto improvement in performance and cost," meaning it claims gains on both fronts at once rather than trading one for the other. The new checkpoint is live on DeepSeek's website, app and API, where it now answers requests sent to the deepseek-flash model name. According to DeepSeek's release note, it was tuned for the company's own DeepSeek Harness agent tool (version v0.2.1-rc.1 or later is recommended), plus Tencent CodeBuddy and WorkBuddy. The weights are on Hugging Face and ModelScope.
It comes just over five weeks after the original V4.1 Flash, which DeepSeek released on September 10, as China's Yicai also noted. The model card calls 1017 "a minor update" with better reasoning, coding, agent and visual scores. DeepSeek says it trained the model with reinforcement learning with self-feedback so it can build whole projects from scratch, with demos ranging from a web racing game to a 3D model of the Forbidden City. The Decoder explains that this method uses the model's own confidence scores or self-critiques as the reward signal, instead of human raters or a separate reward model.
Advertisement
Key details
The benchmark claims
All of the scores below come from DeepSeek's own table, and no independent results have been published yet. On that table, 1017 beats the September V4.1 Flash on all 11 benchmarks listed:
- DeepSWE v1.1 (coding): 77.0, up from 74.2. Claude Opus 5 is listed at 74.0 and GPT-5.6 Sol at 73.0.
- Automation-Bench v1.0.6: 60.5, against 50.3 for Opus 5.
- Agents' Last Exam: 35.4, against 28.6 for Opus 5 and 26.7 for GPT-5.6 Sol.
- Apex Bench v1.1: 72.1, against 65.8 for Opus 5.
- Where Opus 5 still leads: Humanity's Last Exam without tools (56.3 vs 42.6), Terminal-Bench 4.0 (51.8 vs 42.3), ProgramBench (37.0 vs 29.1) and NL2Repo-Bench (75.3 vs 68.2), as MarkTechPost also points out.
The cost-per-task chart
DeepSeek's cost chart plots DeepSWE accuracy against average spend per task. At its top setting, 1017 reaches 77.0% for just under $0.20. Its low setting still clears 72% for about $0.10, according to The Decoder, while the September version scored 74.2% at around $0.50. The Decoder notes that Claude Opus 5, GPT-6 Astra and Gemini 3.8 Flash sit between about $2 and more than $10 per task on the same chart, and none of them match 1017's score. The chart's footnote says rival figures come from the public DeepSWE leaderboard and official OpenAI charts.
Model and API specs
- License: MIT, for both the repository and the weights, per the Hugging Face model card.
- Size: DeepSeek's September note describes V4.1 Flash as a 552B-parameter mixture-of-experts model that activates only 8B parameters for input and 16B for output. Hugging Face's file listing for the 1017 weights shows 763B parameters, and the model card doesn't explain the gap.
- Context: 1M tokens, with up to 384K output tokens on the API, according to DeepSeek's pricing page.
- Price: $0.30 per million input tokens (cache miss) and $1.20 per million output tokens at peak times, with off-peak rates half that. Weekends count as off-peak.
- Other platforms: Vercel added the model to its AI Gateway on the same day, and says it passes provider pricing through with no markup.
Search and Fetch, explained
According to DeepSeek's developer guide, the Search API returns between 1 and 50 ranked results for a query and supports operators such as site:, lang:, after: and before:. Each result includes a relevance score and an "authority" tier. The Fetch API takes one URL and returns the page as Markdown. You're only billed for successful requests, so empty searches and failures cost nothing. Each API is capped at 5 requests per second per account. DeepSeek says the same engine already powers search on its website and API and handles hundreds of millions of queries a day. In its Chinese-language release note, it adds that the engine filters out SEO cheating, GEO (generative engine optimization) manipulation and low-quality AI-generated content. Yicai reported the same claim, along with the China price of 5 yuan per 1,000 calls.
Why it matters
- Cost per task is the new battleground. Agents run many steps, so shorter outputs add up. DeepSeek's pitch is that the price of a whole finished task now matters more than the price per token.
- Open weights keep pressure on closed labs. An MIT license lets companies download the model, run it themselves and change it. Rival open-weight releases keep arriving too, as we covered in Reflection Beam Is Out: 501B Open Weights, Apache 2.0.
- A model maker is now selling search. Developers can now get model calls and web search from one DeepSeek account. If its filtering claims hold up, sites built to game search rankings could become less visible to AI agents.
- The claims are still self-reported. The Decoder notes that independent benchmarks haven't come out yet, and developers' real-world impressions often differ from benchmark tables.
What's next
The free window is the nearest date to watch. DeepSeek API users who reach Search and Fetch through server-side tools in DeepSeek Harness, Claude Code or Codex pay nothing for those calls for now. Billing starts at 16:00 UTC on October 31, 2026, which is midnight on November 1 in Beijing. That free window covers only those server-side tool calls. DeepSeek's docs list direct Search and Fetch requests at the standard per-request price.
Independent testing is the next big question. Outside leaderboards and developers' own projects will show whether the 77.0% DeepSWE score and the cost figures hold up beyond DeepSeek's chosen harness, Mini-SWE-V2. DeepSeek's September note also referred to an upcoming V4.1 Pro model, without giving a date.
FAQ
Is DeepSeek V4.1 Flash 1017 open source?
Yes, the weights are open. The Hugging Face model card says the repository and the model weights are licensed under MIT. The release ships without a Jinja chat template. Instead, DeepSeek provides a Python reference encoder and a Rust toolkit called deepseek-recipe for formatting prompts.
How much do DeepSeek's Search and Fetch APIs cost?
Each costs $0.75 per 1,000 successful requests, or 5 yuan in China, and failed or empty calls aren't billed. Calls made through server-side tools in DeepSeek Harness, Claude Code and Codex are free until 16:00 UTC on October 31, 2026.
Does it really beat Claude Opus 5 and GPT-5.6 Sol?
Only on some tests, and only in DeepSeek's own numbers. It leads on DeepSWE, Automation-Bench, Agents' Last Exam and Apex Bench. Opus 5 still leads on Humanity's Last Exam, Terminal-Bench 4.0, ProgramBench and NL2Repo-Bench. Independent results haven't been published yet.
Sources
- DeepSeek on X – DeepSeek-V4.1-Flash-1017 is here! (thread 1/4–4/4)
- DeepSeek API Docs – DeepSeek-V4.1-Flash-1017: Wish-Driven Programming Made Real
- DeepSeek API Docs (Chinese) – DeepSeek-V4.1-Flash-1017 release note
- DeepSeek API Docs – Models & Pricing
- DeepSeek API Docs – Search & Fetch
- DeepSeek API Docs – DeepSeek-V4.1-Flash: Smarter, Faster, More Efficient
- Hugging Face – deepseek-ai/DeepSeek-V4.1-Flash-1017 model card
- The Decoder – Deepseek V4.1 Flash 1017 beats larger rivals in agentic tasks while once again massively cutting per-task costs
- MarkTechPost – DeepSeek Just Released V4.1-Flash-1017, Scores 77.0 on DeepSWE, Opens Search API
- Vercel – DeepSeek V4.1 Flash 1017 now available on AI Gateway
- Yicai – DeepSeek更新模型了,还对外开放了搜索引擎 (DeepSeek updates its model and opens its search engine)

Comments
Post a Comment