DeepSeek V4.1 Flash 1017: what the new release scores, costs and changes in the API

DeepSeek V4.1 Flash 1017 is a minor update to the company's Flash model, released on October 17, 2026 under the MIT License, with a 1M-token context window and image input. The deepseek-flash API name now routes to it. DeepSeek's model card shows gains on every benchmark it lists. Those numbers are DeepSeek's own and no independent run was available when this was written.

What DeepSeek actually shipped

The release notes list the date as October 17, 2026 and describe stronger benchmark performance, faster task completion and shorter outputs than the previous version. DeepSeek summarizes that as "a Pareto improvement in performance and cost." The Hugging Face model card calls it a minor update to V4.1 Flash with "substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks."

The card lists 763B parameters across mixed tensor types and does not state an active parameter count. That matters, because cost and speed depend on the active count in a mixture-of-experts model, as our explainer on total versus active parameters lays out. The weights are open under MIT, the most permissive of the common licenses, and the card points to vLLM, SGLang and GGUF quantizations for local use.

The release notes say the model is tuned for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy, and advise upgrading to DeepSeek Harness v0.2.1-rc.1 or later. DeepSeek's examples include a 3D model of the Forbidden City, a browser racing game and a recreation of Portal 2 in the Godot engine. Those are demos. They show the model can build a project end to end, not how often it succeeds.

The benchmark gains, and what they do not show

The model card compares the new release with the previous Flash on five tests:

• Terminal-Bench 4.0: 42.3, up from 31.2
• DeepSWE v1.1: 77.0, up from 74.2
• ProgramBench: 29.1, up from 20.3
• Humanity's Last Exam without tools: 42.6, up from 36.8
• Automation-Bench v1.0.6: 60.5, up from 54.8

The biggest jump is on Terminal-Bench 4.0, a gain of 11.1 points. The smallest is DeepSWE, which was already high at 74.2.

The card also sets the model against K3, GLM-5.3, GPT-5.6 Sol and Opus 5. Opus 5 scores higher on most listed benchmarks, while the DeepSeek model leads K3, GLM-5.3 and GPT-5.6 Sol on several. That is a fair summary of where open weights sit: near the frontier on coding and agent tasks, still behind the best closed model on most of them.

The caveats are the usual ones. Every figure comes from the vendor. The release notes show a DeepSWE chart of pass rate against cost per task, run with the mini-swe-v2 harness, but the chart is an image and the page gives no scores in text. Benchmarks also move with the harness, the reasoning setting and the prompt. Our piece on what AI leaderboards rank and what a top score hides covers why a single table rarely settles a purchase.

The API changes that affect existing code

For anyone already calling the API, nothing breaks. DeepSeek's documentation says deepseek-flash "is now served by DeepSeek-V4.1-Flash-1017" and the calling method is unchanged. The older names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted, but their models are retired. Requests to them now go to the new model and are billed at the Flash price. The API follows both the OpenAI and Anthropic formats and exposes a thinking parameter and a reasoning effort setting.

The practical consequence is silent model drift. If your tests were tuned to the old Flash, your outputs changed on October 17 without a code change. Pin evaluations to a dated model where you can, and rerun your own suite.

Pricing, including the peak-hour rule

DeepSeek's pricing page, read on October 19, 2026, lists prices per million tokens with a peak and off-peak rate. Off-peak is half of peak. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday, excluding Chinese public holidays.

• deepseek-flash, cache-miss input: $0.15 off-peak, $0.30 peak
• deepseek-flash, cache-hit input: $0.003 off-peak, $0.006 peak
• deepseek-flash, output: $0.60 off-peak, $1.20 peak
• deepseek-v4-pro, cache-miss input: $0.66 off-peak, $1.32 peak
• deepseek-v4-pro, output: $1.98 off-peak, $3.96 peak

Both models list a 1M-token context. Flash has a concurrency limit of 2,500 and Pro of 500. The page carries no update date, so check it before budgeting. For comparison with US labs, see our frontier API price tracker.

Search and Fetch APIs open to everyone

The release also opens DeepSeek's Search and Fetch APIs, powered by its own search engine, to all developers at $0.75 per 1,000 calls per endpoint. Calls made through server-side tools in DeepSeek Harness, Claude Code and Codex are free for DeepSeek API users for now. Billing starts October 31, 2026 at 16:00 UTC. If you run agents that search heavily, that date is the one to calendar, because a loop that makes thousands of calls will move from free to a line item overnight.

Our take

This is a good release for anyone who wants a cheap, permissively licensed model for coding agents. The MIT license and the price put it in a different class from closed frontier models. The claims are narrower than the framing. Gains of this size on a self-run benchmark table are a reason to test, not a ranking. Run it against your own repository and failure cases, compare cost per solved task rather than cost per token, and watch for independent evaluations before moving production traffic. Also note that the model is hosted in China under DeepSeek's terms, so teams with data residency or procurement rules should use the open weights on their own infrastructure rather than the hosted API.