StepFun introduces Step 5 Preview for agentic work

StepFun presents Step 5 Preview as a model for tasks that involve more than answering a single prompt.[1] Its model card says it can navigate large codebases, coordinate changes across files, run and verify code, and iterate on results.[2] For research and document work, StepFun says it can combine research, multimodal understanding and tools to analyze documents and media, then produce structured reports and analyses.[2] The company identifies finance as a particular strength when tasks combine quantitative reasoning with financial data and supporting documents.[2] These are vendor descriptions of intended capabilities, not guarantees for a particular workflow.

StepFun says Step 5 Preview has 600 billion total parameters, of which 27 billion are active per token, and supports a 1-million-token context window.[1] It also accepts vision input.[1] Its documentation describes an architecture of 92 transformer blocks, alternating Gated Sliding Window Attention with sparse full-attention layers.[5] StepFun says Sparse GQA is intended to support efficient long-context training and inference up to that context length.[5] One operational boundary matters: the model itself does not access a user’s local environment or external services; those capabilities come from the application that integrates it.[3]

StepFun says the model is available through its products and API, and can be served through an OpenAI-compatible API using vLLM or SGLang, or used for offline generation with Transformers.[2][1] The company plans to release open weights on October 15, 2026.[1][5] Artificial Analysis lists input pricing of $1 per million tokens and output pricing of $2.70 per million tokens.[4] Output-limit documentation is inconsistent: StepFun’s platform documentation lists a maximum of 64,000 tokens, while its Hugging Face model card lists up to 1 million tokens.[3][2] Teams should verify the applicable limit for the version and interface they plan to use.

StepFun says Step 5 Preview scored 44 on the Artificial Analysis Intelligence Index.[1] It also reports that approximately 70% of evaluation participants judged the model capable of autonomously solving coding tasks of moderately high complexity.[1] Artificial Analysis estimates an average evaluation cost of $1.03 per task on its Intelligence Index.[4] It measured generation at 89 tokens per second and described that as faster than its 78-token-per-second average.[4] These figures provide evaluation and speed reference points, but do not by themselves establish performance on a team’s own tasks.