StepFun opens Step 5 Preview API at $1 per million input tokens

StepFun, the Shanghai AI lab, has opened API access to Step 5 Preview, a 600-billion-parameter model priced at $1 per million input tokens and $2.70 per million output tokens. StepFun's own materials put the release of the weights on October 15, according to the OrcaRouter blog, which tracked the launch on September 20.

On Artificial Analysis's Intelligence Index, an independent composite of benchmark results, the model scores 44. That ranks it 24th of 200 models, well above the median score of 24.

What developers get today

StepFun's documentation lists the model as step-5-preview. It takes text, images and video as input, up to 60 images per request and video files under 128 MB, and returns text. The context window is 1 million tokens. It supports tool calling, JSON mode and JSON Schema output, prompt caching, streaming, and a reasoning-effort setting with low, medium and high levels.

In plain terms, this is a model built for agents, meaning software that calls tools in a loop, reads large codebases and runs for a long time. The documentation itself stresses programming and professional knowledge work.

The architecture is a sparse mixture of experts. Of the 600 billion parameters, about 27 billion are active for any given token, which is how a model this large can be served at this price. Artificial Analysis measured output at 99.8 tokens per second, with 2.96 seconds to the first token.

The price, and the fine print on it

A dollar in and $2.70 out is cheap for this tier. OrcaRouter reports StepFun's own claim that a single task costs one-eighth of what it does on Claude Opus 5. Artificial Analysis also lists a 95 per cent discount on cached input, which Runtimewire points out could matter for agents that reread the same codebase again and again.

Here is what the headline price hides. Artificial Analysis calls the model "very verbose". It generated 160 million output tokens to complete the index evaluation, against a median of 92 million. You pay per token, so a model that thinks at length eats into its own discount.

Test it on your own workload before moving anything.

How good is it?

StepFun's published numbers are mixed. On Terminal-Bench v4, which tests command-line tasks, Step 5 Preview scores 33.3 against 57.9 for GPT-6 Astra and 52.3 for Claude Opus 5. On FrontierFinance it scores 66.4, ahead of Astra's 55 and behind Opus 5's 69.7. On DRACO it posts 83.3, again between the two. These are vendor-run evaluations, and OrcaRouter notes they have not been reproduced by third parties.

The independent picture is simpler. A score of 44 puts it level with Kimi K3 and a point behind GLM-5.3, by OrcaRouter's reading of the index. That is a good result, and it is short of the frontier.

What is promised but not delivered

The weights are the open question. A Hugging Face repository named stepfun-ai/Step-5-Preview-BF16 was created on September 20, and when OrcaRouter checked it held a single placeholder file, with no weights, licence or model card. StepFun has released open weights before. Its Step 3.5 Flash, a 196-billion-parameter model, came out on February 12 this year. But the licence for Step 5 has not been disclosed, and the licence decides whether a business can use the weights at all.

There are loose ends on the specifications too. OrcaRouter found a third-party configuration listing a 350,000-token context and a 64,000-token output cap, which conflicts with the 1 million in StepFun's documentation. Even the launch date is fuzzy. Artificial Analysis records the release as September 18, while press coverage dates the announcement to September 20.

Mark October 15. If the weights arrive under a permissive licence, a model at this level becomes something companies can run on their own servers. If they arrive late or restricted, Step 5 Preview is one more cheap API among many.