DeepSeek's New V4.1-Flash Holds Prices Steady While Trimming Cost per Task
DeepSeek has pushed a model with better performance and faster task handling to its web, app, and API channels. Per-token pricing is unchanged, but the company says shorter answers mean each job costs less.
DeepSeek released DeepSeek-V4.1-Flash-1017 on October 17. The update has been rolled out across the web interface, the mobile app, and the API, the channel developers use to plug the model into their own products. The model weights are also published under an MIT license, so anyone can download and run it. The company says benchmark scores have risen compared with the previous version, while the time needed to finish a task and the length of its answers have both gone down.
What got better
This update focuses on agentic work, meaning the model’s ability to handle a job on its own across multiple steps. DeepSeek says it fed the model’s own reflections on its work back into training as feedback. That approach, the company claims, improved its ability to carry out project development work such as building games and 3D modeling.
The benchmark results DeepSeek published show gains over the earlier V4.1-Flash. Terminal-Bench 4.0 climbed from 31.2 to 42.3. DeepSWE v1.1 rose from 74.2 to 77.0, and ProgramBench went from 20.3 to 29.1.
Does it beat the competition?
The answer varies by test. In DeepSeek’s own comparison table, some agentic evaluation scores came in above GPT-5.6 Sol and Opus 5, while others fell below them. Terminal-Bench 4.0 and DeepSWE v1.1 were each measured using a specified execution harness. Some HLE scores were measured on text-only questions rather than the full problem set.
Same rate, lower bill
The previous V4.1-Flash, released on September 10, led with an API price cut. This time the per-token rate stays put and the gains come from task performance and speed instead. DeepSeek says that because task completion time and output length have both shrunk, the cost of finishing a single job is lower.
API charges are calculated in tokens, the small text fragments models read and write. During off-peak hours, the rate is $0.15 per million tokens for fresh input, $0.003 when reusing input that was cached earlier, and $0.60 for output. During peak hours, each rate doubles. When answers get shorter, the burden drops for users who pay by volume of output.
Where you can use it
Regular users get the new version right away on the web and in the app. On the API, it keeps the existing model name, deepseek-flash. Services already wired up to that name will pick up the new version without any change. Users who want to run the model themselves can download the released weights.
DeepSeek has also opened up its in-house search engine through Search and Fetch APIs. The price is $0.75 per 1,000 calls. The company announced that Search and Fetch, when called as server-side tools from DeepSeek Harness, Claude Code, and Codex, are free for DeepSeek API users. Billing begins October 31 at 16:00 UTC.
The gist of DeepSeek’s pitch is that you get faster, shorter task completion at the same rate. For real-world use, it is worth measuring both completion time and cost together on the jobs you actually hand off. For users in Korea, the open questions are whether the service is accessible domestically and how well it handles Korean-language work.