StepStar has open-sourced Step Code, which matches DeepSeek Harness in self-test performance and delivers lower token consumption.

Beating AI News Flash: Chinese AI startup StepFun has open-sourced its AI coding tool Step Code v0.1.0 under the MIT license. The tool runs directly in the terminal, supporting code reading and writing, test execution, and development task handling, with compatibility for MCP, Agent Skills, and Claude Code plugins. Its built-in StepPage feature allows users to publish local static websites online with a single command. Step Code’s core highlight is low token consumption. In StepFun’s self-conducted Terminal-Bench 2.1 benchmark, it completed 72 out of 89 tasks, scoring 80.9% – tying DeepSeek Harness for the highest score while using fewer tokens. In StepFun’s in-house Multi-Frame long-task test (150 questions total), Step Code finished 110 tasks for a 73.3% success rate, averaging 5.09 million tokens per task. Among the 6 tested harnesses, it ranked first for both highest task completion rate and lowest token usage. However, StepFun did not disclose the specific underlying model used for the benchmark, nor did it detail how token consumption was reduced. The company only noted that Step Code was optimized for token efficiency alongside its Step series models. Numerous token-saving designs are visible in the source code: file reads and command outputs are length-limited, search results are narrowed as much as possible, long conversations default to context compression, and the system prompt prioritizes direct tool calls over spawning additional agents. There is also a more aggressive contextProjection feature that trims old tool outputs, duplicate results, and historical reasoning, though it is disabled by default. Overall, Step Code’s token advantage comes from multiple optimizations to both the model and the test harness.