About twenty cents. Deepseek’s own chart puts the cost of its new Flash model completing a single task on the DeepSWE coding benchmark at roughly that figure. On the same chart, competing models from OpenAI, Google and Anthropic range from $2 to more than $10 per task.
That figure is the main selling point of Deepseek-V4.1-Flash-1017, a refreshed version of the Chinese AI firm’s Flash model. You can use it now through Deepseek’s website, app and API. Deepseek has also released the weights as open source on Hugging Face and ModelScope, which means you can download the model and run it locally.
The savings come from shorter answers
Most of the model’s cost edge comes from how much text it generates. According to Deepseek, the 1017 release writes shorter outputs and completes tasks more quickly than the model it replaces, V4.1-Flash, and it also scores higher on benchmarks. The company describes this as a Pareto improvement in the performance-cost trade-off. Put simply, it improved on one measure without slipping on the other.
The DeepSWE chart backs this up with figures. On its highest setting, V4.1-Flash-1017 reaches 77.0 percent at an average of just under $0.20 per task. On the low setting it still scores above 72 percent, for around $0.10.
The older model scored 74.2 percent at roughly $0.50 per run. That gives the new version close to three extra points of accuracy at well under half the cost per task. If your agents work through hundreds of tasks every day, that difference grows fast.
Larger rivals score lower, at least on Deepseek’s charts
Deepseek’s benchmarks show V4.1-Flash-1017 ahead of bigger competitors such as Claude Opus 5, GPT 5.6-Sol, Kimi K3 and GLM 5.3 in several coding and AI agent categories. On the DeepSWE cost chart, GPT-6 Astra, Gemini 3.8 Flash and Claude Opus 5 each cost more per task, and none of them reaches the accuracy of the smaller Deepseek model.
Treat these results with some caution. The benchmarks come from the vendor, and no independent results are out yet. Benchmarks also cover only part of the picture. Developers can judge code quality very differently from how a leaderboard ranks it, and a model that does well on a test suite may still produce code you would not want to maintain.
Deepseek optimized the model for particular development environments: its own Deepseek Harness (with version v0.2.1-rc.1 or later recommended), Tencent CodeBuddy and WorkBuddy. Don’t expect the same numbers if you run it in another tool.
The model marks its own work
To build complete software projects from scratch, Deepseek uses a post-training technique called reinforcement learning with self-feedback, or RLSF. The model’s own internal confidence scores or self-written critiques act as the reward signal, so no human evaluators or external reward model are needed.
Deepseek says this makes end-to-end project development better, covering everything from game programming to 3D modeling and design. The model can also animate short explainer videos, much like the recently released Opus 5.5.
Having a model reward itself keeps costs low. Whether that confidence matches output that actually works is just what independent testing will have to find out.
Developers get access to Deepseek’s search engine
For people building agents, the less eye-catching announcement may matter more. For the first time, Deepseek is making its web search tools, the Search & Fetch APIs, available to every developer. The company says they run on a search engine it built in-house, which already processes hundreds of millions of queries per day around the world.
Each endpoint is priced at $0.75 per 1,000 calls. Deepseek API users can use the features at no cost through server-side tools in Deepseek Harness, Claude Code and Codex until October 31, 2026 (4:00 PM UTC). Deepseek’s API documentation explains how to set them up.
The real risk isn’t the model itself
Relying heavily on Chinese AI models in your workflow became a bit less attractive this week. The Chinese government is reportedly planning stricter rules on publishing open-weight models, and might even stop their release altogether. One factor in the decision is whether a system ranks among the top 10 most capable models.
A model that outperforms Claude Opus 5 on coding benchmarks is precisely the kind that rule could catch. If you plan to build on V4.1-Flash-1017, grab the weights from Hugging Face while they’re still available, and keep a tested fallback model ready in case future releases stop arriving.