DeepSeek
DeepSeek upgrades V4.1 Flash with higher scores and no price change
DeepSeek replaced V4.1 Flash with a 1017 build that it says scores higher and uses fewer tokens per task, at the same per-token prices.
DeepSeek replaced its V4.1 Flash model with a new build, DeepSeek-V4.1-Flash-1017, and left per-token prices where they were. The company says the new version scores higher on every benchmark it compared and uses less time and fewer output tokens per task.
DeepSeek announced the upgrade in an API changelog entry dated Oct. 17, 2026. The entry says the official web and app models were updated too, and that “simply use deepseek-flash to access the latest model.” The calling method does not change.
Scores
DeepSeek’s own testing shows gains on all nine benchmarks it ran for both the new build and the original V4.1-Flash, which the company released Sept. 10, 2026. None of the figures is an independent result.
The largest jumps came on agent and coding tasks. Terminal-Bench 4.0, run with the Claude Code framework, rose to 42.3 from 31.2. ProgramBench rose to 29.1 from 20.3. Humanity’s Last Exam without tools rose to 42.6 from 36.8, or 44.0 from 39.1 on the text-only subset.
Smaller gains came on DeepSWE v1.1 (77.0, up from 74.2), NL2Repo-Bench (68.2, from 65.4) and Agents’ Last Exam (35.4, from 31.8). With tools enabled, Chartography reached 83.0 from 78.9, BabyVision 92.5 from 89.6 and ZeroBench-main 51.0 from 49.0 at Pass@5. DeepSeek also reported 60.5 on Automation-Bench v1.0.6 and 72.1 on Apex Bench v1.1, with no September score to compare against.
DeepSeek calls the result a “performance–cost Pareto improvement” and says the shorter runtimes and outputs let users finish tasks at lower cost and higher speed. The company recommends DeepSeek Harness 0.2.1-rc.1 or later, and says it tuned the model for Harness, Tencent CodeBuddy, WorkBuddy and other platforms.
Pricing
Per-token rates are the same as at launch. DeepSeek’s pricing page, read Oct. 17, 2026, lists deepseek-flash at $0.15 per million input tokens and $0.60 per million output tokens off-peak, and $0.30 and $1.20 at peak. Cache-hit input costs $0.003 off-peak and $0.006 at peak. Peak hours are 01:00-04:00 and 06:00-10:00 UTC Monday through Friday, excluding Chinese public holidays.
The model keeps a 1 million-token context window, a 384,000-token maximum output, thinking and non-thinking modes, and vision support.
The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work. Requests to them are now served by the 1017 build at the Flash price, where the page previously said V4.1-Flash.
The changelog entry also opens DeepSeek’s in-house Search API. The pricing page lists Search and Fetch at $0.75 per 1,000 successful requests, free through DeepSeek Harness, Claude Code or Codex until per-request billing starts at 00:00 Beijing time Nov. 1, 2026.
Our View
DeepSeek cut output length and left the price per output token alone. Output carries the highest rate on the price list, $0.60 to $1.20 per million tokens. A task that produces fewer tokens costs less to run.
Developers get this without any work. The deepseek-flash name and the legacy names point at the new build. The model changes without anyone choosing it. Anyone who tuned prompts against the September build should rerun their own tests.
The lower bill depends on the shorter outputs holding up on a team’s own workload.