One racing-game generation task consumed roughly 60 million tokens, took 51 minutes, and cost about 3 yuan. The significance of DeepSeek-V4.1-Flash-1017 goes beyond its low price: it is beginning to optimize the number of tokens and tool calls an Agent needs to complete an entire task.
What this is
DeepSeek released V4.1-Flash-1017 on October 17, with a focus on improving its Terminal-Bench, Automation-Bench, and ProgramBench results, covering terminal operations, tool execution, and programming tasks. The company also introduced a “performance-cost Pareto improvement”: unit prices remain unchanged, but the total cost of delivering real requirements falls through fewer invalid outputs and shorter task chains.
The author used DeepSeek Harness, the model’s execution environment and tool-calling framework, to recreate a racing game. In the first round, the author supplied only a one-sentence brief, and the game could already support driving, lap counting, and camera switching. In the second round, the author added visual and interaction requirements, substantially improving the result. The task ran two projects simultaneously, consumed roughly 60 million tokens, and cost about 3 yuan.
Industry view
We believe this matters more than model benchmark scores in isolation. The cost of an Agent—an artificial intelligence system that can autonomously break down tasks and invoke tools to execute them—comes not only from the unit price of tokens, but also from repeated requests, tool calls, correction loops, and expanding context. If vendors can reduce the total resources required to complete a task, enterprises will find it easier to move automation from demo projects into stable workflows.
The risks are equally clear. Roughly 60 million tokens and 51 minutes still mean that generating the game required substantial resources, and the cost of rework on complex projects is not included. Most official benchmarks come from vendors, so they do not yet prove that costs can be reduced reliably in every setting. A racing game also has clear visual requirements and rapid feedback, so these results cannot be extrapolated directly to data analysis, customer service, or enterprise system modernization. Connecting sensitive data, verifying outputs, and ensuring reliability on long-running tasks remain deployment barriers.
Impact on regular people
For enterprise IT: When procurement teams evaluate models, they should ask less about “How much does each call cost?” and more about “How much does it cost to complete a real workflow, how long does it take, and how many human interventions are required?” This will further lower the cost threshold for small-batch software prototypes and internal-tool automation.
For individual careers: Nontechnical employees can increasingly use natural language to generate small, runnable tools, making prompting, material preparation, and acceptance testing more important. In the short term, this lowers the barrier to experimentation rather than replacing entire roles.
For consumer markets: The supply of demand-driven custom websites and mini-games will increase, and prices may fall. At the same time, products with similar functionality but little capacity for maintenance will face more pressure from model-based generation tools.