DeepSeek-V4.1-Flash-1017: what's new and what it costs

DeepSeek-V4.1-Flash-1017 is an update to DeepSeek's V4.1 Flash model, released October 17, that scores higher on coding and agent tests and writes shorter answers. The weights are free under the MIT license, and the API costs $0.30 per million input tokens and $1.20 per million output tokens at peak hours, half that off-peak.

DeepSeek announced it on X. @deepseek_ai promised "stronger benchmark performance, faster task completion, and shorter outputs" than the last version and called it "a Pareto improvement in performance and cost." Translated: it got better and cheaper to run at the same time. It's live on DeepSeek's website, its app and the API.

It also leaked a few hours early. @aitrackerbot caught the name deepseek-v4.1-flash-1017 in DeepSeek's API model list the night before, and then it disappeared. The "1017" is a date stamp, the same naming DeepSeek used for earlier snapshots like V4-Flash-0731.

What changed in V4.1-Flash-1017

The model card on Hugging Face calls it "a minor update to DeepSeek-V4.1-Flash." The numbers moved more than that suggests. These are DeepSeek's own scores:

Terminal-Bench, where a model works in a command line to finish real tasks, went up 11 points. DeepSeek ran that test inside the Claude Code harness. Its table also has 1017 beating Kimi K3, GLM-5.3 and GPT-5.6 Sol on most of the coding and agent tests. Opus 5 still leads on hard reasoning (HLE) and on terminal work.

The second post in DeepSeek's thread says training used "reinforcement learning with self-feedback" so the model could build whole projects on its own. The release notes show it making a browser racing game, a 3D model of the Forbidden City, and a Portal 2 remake in the Godot engine.

The model card lists 763 billion total parameters and a 1 million token context window. DeepSeek doesn't say how many parameters are active per token.

What V4.1-Flash-1017 costs

You can use it three ways:

• Download the weights. They're free under MIT, which allows commercial use. The model card links quantized builds for llama.cpp, Ollama and LM Studio, but a 763B model still needs a big server to run.
• Call the API. DeepSeek's pricing page lists $0.30 per million input tokens on a cache miss, $0.006 on a cache hit, and $1.20 per million output tokens at peak. Off-peak is half price. Peak runs 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, which is evening and overnight in Chicago, so most US business hours get the discount.
• Chat with it on DeepSeek's website or app.

Existing API users don't have to change anything. The deepseek-flash model name now points to 1017, and the retired deepseek-v4-flash name routes there too, billed at the Flash price.

DeepSeek also opened its Search and Fetch APIs to everyone at $0.75 per 1,000 calls. When they run as server-side tools in DeepSeek Harness, Claude Code or Codex, they're free for API users until October 31.

My read

I'd watch the shorter outputs most. You pay for output tokens, so a model that gets to the answer in fewer words costs less even when the per-token price doesn't change. Combine that with Terminal-Bench scores closing in on the big closed models, and a long, boring agent job keeps getting cheaper to hand off.

The scores come from DeepSeek's own table, though, and I couldn't find independent results yet. And DeepSeek's hosted API sends your prompts to a company based in China, so anything touching client records needs a data decision before anyone tests it. Running the MIT weights yourself, or waiting for a US inference provider to host the model, gets around that.

What it means for a business

A cheap agent model is only useful once you know which jobs to give it, what data it can see and who checks its work. New Face Design's free process audit goes through your week with you to find the repetitive work worth automating and the data that should stay on your own systems.