DeepSeek has released V4.1 Flash 1017, an open-weight update to its Flash model line, five weeks after the September version, and it is selling the release on cost: the company says the model completes agent tasks for just under $0.20 each, down from about $0.50.
The weights are published under the MIT licence. On DeepSeek's own DeepSWE chart, the update scores 77.0 per cent; in the company's comparison table it beats Anthropic's Claude Opus 5 on four of eleven rows — though notably against Opus 5 rather than the newer Opus 5.5, and independent benchmarks have yet to confirm the figures.
Developers calling the deepseek-flash endpoint on DeepSeek's API are already on the 1017 snapshot, the company says, which means September-era test results may no longer describe the model actually serving traffic. Builders are being advised to rerun their own agent and coding evaluations before relying on older numbers.
The release underlines the price war reshaping the AI market: open-weight models that can run coding and agent workloads at a fraction of frontier-lab prices are pressuring the subscription and API economics of rivals in the United States and Europe.
