StepFun Step 5 Preview Hits OpenRouter at $1 Input, Weights Due Oct 15

StepFun Step 5 is live on OpenRouter, and coders are poking at it before the big open-weights date. The Shanghai lab's 600-billion-parameter flagship sits behind OpenRouter's common API at $1 per million input tokens, and StepFun says open weights follow on October 15. The pitch is agentic coding and finance work at a lower price than frontier rivals.

What actually happened

On October 8, the OpenRouter listing for StepFun Step 5 Preview went live. OpenRouter describes it as StepFun's flagship for agentic work, built on sparse MoE with 27B active out of 600B total parameters. It says the model is strong in software engineering and professional knowledge work, with particular strength in finance, and built for long tasks across big codebases and documents.

One detail matters for reliability: the listing shows a single provider. Every request goes straight to StepFun, so there is no fallback host if it has a bad day.

Vercel also added Step 5 Preview to its AI Gateway. Its changelog says the model takes text and image input with a 1M-token context window and frames it for coding, knowledge work and financial analysis. Vercel says its gateway passes through provider pricing with no markup.

Don't call it a brand-new model, though. Bad Signal dates StepFun's own API announcement to September 20, so October 8 is the OpenRouter listing date, not the debut.

The numbers that matter

Bad Signal summarizes the company pitch: a 600B MoE that activates 27B parameters per token, a 1M-token context, and an "API now, weights later" split. It also cites an independent Artificial Analysis card with an Intelligence Index score of 44, output speed of about 99.8 tokens per second and roughly $0.71 per Intelligence Index task.

On price, Step 5 sits well above StepFun's own budget lineup. Step 3.5 Flash lists at $0.10 input and $0.30 output per million tokens, and Step 3.7 Flash at $0.20 and $1.15. Step 5 Preview is $1.00 and $2.70. Caching is the big lever: cache reads cost $0.05 per million, a 95% cut from uncached input.

Benchmarks are mixed. OpenRouter shows Vals AI scores including 70.4% on Vibe Code Bench v1.1, 50.7% on Finance Agent v2, 31.8% on Terminal-Bench 4.0 and 49.4% on the Vals Index. Marketing claims, like AIBase's line that a single task costs about one-eighth of Claude Opus 5, are StepFun-side framing and not independent measurements.

Translation: decent coder, solid finance analyst, not an Opus killer.

Why StepFun Step 5 is worth a look

Chinese labs keep pushing big MoE models into Western developer pipes, and OpenRouter plus Vercel is about as wide a distribution channel as it gets. We saw the same playbook with Kimi K2.6. If StepFun ships the weights on October 15 as promised, Step 5 becomes a self-hostable option for teams that want an agentic model without sending code to a US lab.

The open-weights date is the countdown to watch. StepFun already publishes earlier models on Hugging Face, including several Step 3.5 Flash and Step 3.7 Flash variants, so the channel exists. But a scheduled release is not a release. Until a repo, a license and deployment configs actually show up, self-hosting plans are just plans.

There is also the bill. A cheap sticker price can hide a chatty model, and long agent loops burn output tokens fast. Cost per finished task is the number that matters, not cost per token.

What to do as a reader

If you build with agents, test the model on your own workload before switching anything. Run the same tasks you already give your current model, track cost per finished task and watch how much output it produces.

Price is only half the story. Because the OpenRouter listing names a single provider, a busy day at StepFun means slower responses or errors, with nowhere to reroute. If uptime matters for your product, keep a second model wired in as a fallback.

Check StepFun's Hugging Face page on October 15 before planning any self-hosting, and read the license before you build on it. For the bigger model race, see our year-end AI model odds breakdown. Nothing here is investment advice, just tooling advice.