DeepSeek has quietly replaced the model behind its cheapest tier. On 17 October 2026 it swapped DeepSeek-V4.1-Flash for a refreshed checkpoint, DeepSeek-V4.1-Flash-1017, across its website, app and API, and put the weights on Hugging Face under the MIT licence. The pitch is a better trade between quality and cost: higher scores on DeepSeek's own tests, while finishing tasks sooner and writing less. For anyone already paying for Flash, the change arrived without any action on their part.
DeepSeek's changelog lists gains on all eleven benchmarks it reports. The largest are on agentic coding: Terminal-Bench 4.0 rose from 31.2 to 42.3, run inside the Claude Code harness, and ProgramBench from 20.3 to 29.1. Humanity's Last Exam without tools went from 36.8 to 42.6, and DeepSWE v1.1 from 74.2 to 77.0. The company says it used reinforcement learning with self-feedback on end-to-end projects and tuned the model for its own DeepSeek Harness, recommending version 0.2.1-rc.1 or later, as well as Tencent's CodeBuddy and WorkBuddy. It gives no figure for how much shorter outputs are.
Prices stay as they were. API users keep calling deepseek-flash, and the older model names now route to the new checkpoint and are billed as Flash. Our review of DeepSeek-V4.1-Flash lists $0.30 per million input tokens and $1.20 per million output at weekday peak, half that off-peak, so shorter answers translate directly into a smaller bill. The same release opened DeepSeek's in-house search engine to developers through Search and Fetch APIs at $0.75 per 1,000 calls for each endpoint, a useful option for agents that need web access without a separate provider.
The update matters because Flash has become a default for cost-sensitive agent work, both through the DeepSeek API and in self-hosted setups. A dated checkpoint rather than a new version number also signals that DeepSeek is moving to frequent incremental releases, much as other labs ship dated snapshots. Teams that pin a model version should note that the old names no longer point to the old weights, so behaviour may shift without a code change.
The caution is familiar. Every score comes from DeepSeek, and the comparison with the September release is ours, drawn from two separate changelog entries. Some figures use special settings, such as a text-only subset of Humanity's Last Exam. Independent coding results, especially on long real-world tasks, will show whether the gains hold outside DeepSeek's own harnesses. Until they do, test it on your own workload before relying on the improvement, and check whether tighter outputs drop detail you need.