Article
StepFun open-sources Step Code: self-test matches DeepSeek Harness, uses fewer tokens
Beating AI news flash from Dongcha: StepFun has open-sourced the AI programming tool Step Code v0.1.0 under the MIT License. It runs directly in the terminal and can read and write code, run tests, and execute development tasks. It also supports MCP, Agent Skills, and Claude Code plugins. Its built-in StepPage can publish a local static website online with a single command.
Step Code focuses on using fewer tokens. In StepFun’s own Terminal-Bench 2.1 comparison, it completed 72/89 tasks and scored 80.9%, tying with DeepSeek Harness for the highest score shown in the chart while using fewer tokens. StepFun’s self-developed Multi-Frame long-task test contained 150 questions. Step Code completed 110, scored 73.3%, and consumed an average of 5.09 million token per question, achieving both the highest pass rate and the lowest token usage among the 6 harnesses compared.
However, StepFun did not disclose which underlying model was used in this comparison, nor did it break down how these tokens were saved. The official statement only mentioned that Step Code was optimized together with the Step series models for token consumption. The source code shows that it is relatively “stingy with tokens” throughout the harness: file reads and command output are actively length-limited, searches try to return only the necessary portions, long conversations default to context compression, and the system prompt also requires calling tools directly instead of opening an additional Agent whenever possible. It even has a more aggressive context-trimming mechanism that can delete historical tool outputs, duplicate results, and old reasoning processes, though this is disabled by default.