StepFun Open-Sources Step Code: Self-Tests Match DeepSeek Harness, With Lower Token Consumption

Beating AI News Flash: StepFun has open-sourced its AI coding tool Step Code v0.1.0 under the MIT license. It runs directly in the terminal, can read and write code, run tests, and execute development tasks, and also supports MCP, Agent Skills, and Claude Code plugins. The 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 with a score of 80.9%, tied for the highest with DeepSeek Harness, but with lower token usage.

StepFun's self-built Multi-Frame long-task test has a total of 150 questions. Step Code completed 110, scoring 73.3%, with an average of 5.09 million tokens consumed 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 round of comparison, nor did it break down how these tokens were saved. The official statement only mentioned that Step Code and the Step series models were tuned together for token consumption.

Many "throttling" designs can also be seen in the source code: file reads and command outputs are length-limited, searches try to narrow the return scope as much as possible, long conversations compress context by default, and the system prompt also requires that if a tool can be called directly, an extra Agent should not be opened.

There is also a more aggressive contextProjection, which can trim old tool outputs, duplicate results, and historical reasoning, but it is disabled by default.

Overall, its token advantage comes from the combined effect of multiple optimizations in the model and the harness.