DeepSeek V4.1-Flash-1017: What Cheaper Agent Runs Change

By Jaspal Singh October 18, 2026 Updated: October 18, 2026

DeepSeek-V4.1-Flash-1017, released on 17 October 2026, is an open-weight update to DeepSeek's cheapest model that finishes coding and agent tasks with shorter outputs. On DeepSeek's own tests it scores higher than its predecessor at well under half the cost per task. API users get it automatically, with no code changes.

What Happened

On 17 October 2026 DeepSeek published a release note for DeepSeek-V4.1-Flash-1017, describing it as a Pareto improvement in performance and cost: better benchmark scores, faster task completion and shorter answers than the V4.1-Flash model it replaces. The model went live the same day on DeepSeek's website, its app and its API, and the weights were posted to Hugging Face and ModelScope.

The Hugging Face model card lists 763 billion parameters, a 1 million token context window and weights stored in a mix of FP8 and BF16. It is released under the MIT licence, which covers both the repository and the weights. The card itself calls the release "a minor update" to V4.1-Flash, but the benchmark gains are not minor. On DeepSWE v1.1, a software-engineering test, the score rose from 74.2 to 77.0. On Terminal-Bench 4.0 it rose from 31.2 to 42.3, and on ProgramBench from 20.3 to 29.1.

Cost is the headline claim. DeepSeek's chart plots DeepSWE scores against the cost of each task. As The Decoder reported from that chart, the new model at its top setting costs just under $0.20 per task, against about $0.50 for the previous version. A lower-effort setting still scores above 72% at about $0.10 per task. Rival models in the same chart cost between $2 and more than $10 per task.

DeepSeek also opened two tools to every API user: a Search API and a Fetch API, built on its in-house search engine. Each costs $0.75 per 1,000 successful calls. Calls made through server-side tools in DeepSeek Harness, Claude Code and Codex are free until 16:00 UTC on 31 October 2026 (00:00 Beijing time on 1 November).

Why It Matters

For most people the important number is not the benchmark score but the cost per finished job. Agent tools such as Claude Code, Codex and DeepSeek's own Harness do not answer one question and stop. They read files, run commands, read the results and try again, sometimes for dozens of rounds. Every round sends tokens in and pulls tokens out. A model that reaches the same answer with shorter outputs and fewer rounds cuts the bill even if the price per token never changes, and that is what DeepSeek says happened here.

It also matters that this is open weight. Open models that match closed ones on coding work put a ceiling on what the closed labs can charge. Our earlier look at Mistral Large 4 and what a 1T open-weight model changes made the same point from Europe; DeepSeek is making it again at a much lower price. A company that cannot send source code to a US or Chinese cloud can, in principle, run this model on its own servers under a licence that allows commercial use.

Existing DeepSeek API users are affected whether they asked for it or not. According to the DeepSeek pricing page, the old model names deepseek-v4-flash and deepseek-v4-flash-vision-exp now route to V4.1-Flash-1017 and are billed at the Flash rate. Anyone who tuned prompts or parsers around the old model's output style should re-test before relying on it.

The release also lands during a tense period for Chinese AI. A NIST evaluation we covered in May, which found DeepSeek V4 Pro about eight months behind US frontier models, showed the gap narrowing. The Decoder notes reports that Beijing may restrict future open-weight releases, which would make this kind of drop harder to repeat.

How It Works

DeepSeek says the gains come from post-training, not a new base model. The release note says the model was trained with reinforcement learning with self-feedback to improve end-to-end project development. In ordinary reinforcement learning from feedback, a human or a separate reward model scores each attempt and the model learns to produce higher-scoring answers. With self-feedback, the model's own confidence or self-generated critique supplies part of that signal. That is cheaper to scale because it needs fewer human graders. The risk is that a model grading itself can learn to like its own mistakes, and DeepSeek has not yet published how it checks for this.

The cost saving comes mostly from output length. Output tokens are the expensive part of an API bill: DeepSeek charges $1.20 per million output tokens at peak, against $0.30 per million for fresh input and just $0.006 per million for cached input. Agent sessions resend the same project files over and over, so most input hits the cache and costs almost nothing. What remains is the model's own writing, meaning its reasoning, code and tool calls. If a model reaches a working patch in fewer, shorter steps, the expensive part of the bill shrinks.

Timing also matters. DeepSeek halves every price outside its peak hours, which run from 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays. In India those windows fall at 6:30 to 9:30 am and 11:30 am to 3:30 pm. A team running long batch jobs, such as nightly test-fixing runs, can schedule them outside those hours and pay half.

The Search and Fetch tools fill a practical gap. An agent that needs current documentation can call Search, which takes a query of up to 128 characters and returns up to 50 results with titles, snippets, longer summaries and a quality score. It can then call Fetch to turn one chosen page into Markdown. According to the Search and Fetch API guide, only successful calls that return content are billed, and each tool is limited to five requests per second per account.

Running the model yourself is a different matter. Total size is listed as 763 billion parameters, and the card does not state how many are active per token. Even stored in eight-bit form that is hundreds of gigabytes of weights, plus memory for a context that can reach a million tokens. DeepSeek's example uses an SGLang server across all available GPUs. This is a data-centre model. Meta's recent open release, which we explained in running an 816B model on Mac Studios, shows what extreme home setups look like and why they are slow.

If you only want to try it, the simplest routes are DeepSeek's chat app, its API with the model name deepseek-flash, or a reseller such as OpenRouter. DeepSeek recommends a temperature of 1.0 and allowing at least 256,000 output tokens, because long agent runs produce long transcripts.

What's Still Unknown

Every benchmark figure here comes from DeepSeek. No independent lab had published results when this article was written, and vendor charts tend to pick favourable settings. The comparison table also shows clear limits. Claude Opus 5 still leads on Terminal-Bench 4.0 (51.8 against 42.3), ProgramBench (37.0 against 29.1), NL2Repo-Bench (75.3 against 68.2) and the HLE reasoning test without tools (56.3 against 42.6).

The per-task costs are read from a chart, not a table, and they depend on the harness used, in this case mini-swe-v2. Your own costs will vary with project size, cache hit rates and how many retries your agent makes. DeepSeek has also not said how much faster the model is in seconds, only that tasks finish sooner.

Several technical details are missing: the number of active parameters, a full description of the self-feedback training, and how the model behaves on very long contexts in practice. Also unclear is how long the free tier for Search and Fetch inside Claude Code and Codex will matter, since it ends within two weeks. Finally, if reported Chinese government limits on open-weight releases take effect, it is uncertain whether later DeepSeek updates will be published the same way.

Frequently Asked Questions

Is DeepSeek-V4.1-Flash-1017 free to download?

Yes. The weights are published on Hugging Face and ModelScope under the MIT licence, which covers both the code repository and the model weights. That allows commercial use, modification and redistribution. Downloading is free, but running a 763-billion-parameter model locally needs a multi-GPU server; most people will use it through DeepSeek's paid API or a third-party host instead.

Do I need to change my code to use the new model?

Usually not. DeepSeek's pricing page says the API model name is now deepseek-flash, and the older names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted. Requests sent under those legacy names are now answered by V4.1-Flash-1017 and billed at the Flash price. Test your prompts anyway, because a different model can format answers differently.

How much does the DeepSeek Flash API cost?

At peak times DeepSeek lists $0.30 per million uncached input tokens, $0.006 per million cached input tokens and $1.20 per million output tokens. Off-peak rates are exactly half. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, which is 6:30 to 9:30 am and 11:30 am to 3:30 pm in India.

What are DeepSeek's new Search and Fetch APIs?

They are two web tools that agents can call. Search returns structured results from DeepSeek's own search engine, and Fetch turns a single web page into Markdown. Each costs $0.75 per 1,000 successful calls, and failed or empty calls are not billed. Calls made through DeepSeek Harness, Claude Code or Codex are free until 16:00 UTC on 31 October 2026.

Is V4.1-Flash-1017 better than Claude Opus 5?

Not across the board, going by DeepSeek's own table. It beats Opus 5 on DeepSWE v1.1, Automation-Bench, Agents' Last Exam and Apex Bench, but trails it on Terminal-Bench 4.0, ProgramBench, NL2Repo-Bench and the HLE reasoning test. These are vendor-reported numbers, and no independent evaluation had been published when this article was written.

What is reinforcement learning with self-feedback?

It is the training method DeepSeek credits for the update. Instead of relying only on human raters or a separate reward model, the model's own confidence or self-critique is used as part of the training signal. DeepSeek says it used this to improve end-to-end project building, but it has not yet explained how it guards against the model rewarding its own mistakes.

Can I run DeepSeek-V4.1-Flash-1017 on a Mac or gaming PC?

Not realistically at full size. The model card lists 763 billion parameters stored in FP8 and BF16, which means hundreds of gigabytes of weights before you add memory for a long context. DeepSeek's example uses an SGLang server across all available GPUs. Smaller quantised community versions may appear, but expect them to lose some quality.