StepFun's Step 5 Preview matches Grok 4.6 and Kimi K3 on the intelligence index, at $0.71 a task (2026)

StepFun announced Step 5 Preview on September 20, describing it as a flagship model built for agentic work: software engineering, professional knowledge work and finance. 1 The Beijing lab says the model is available today through its products and API, with open weights promised for October 15. 1 2

Artificial Analysis scores Step 5 Preview at 44 on its Intelligence Index v4.3.2, the same score as xAI's Grok 4.6 and Moonshot's Kimi K3, and puts its cost at $0.71 per completed task against $1.86 for Grok 4.6 and $2.00 for Kimi K3 — against a median of 24 in that price band. 3 4

Signal	Confirmed detail	What it means for you
What shipped	A sparse mixture-of-experts model with 600B total parameters, 27B active per token, a 1M-token context window, and text, image and video input. 1	It runs behind StepFun's API under the model ID `step-5-preview`.
What it costs	$1.00 per 1M input tokens, $0.05 for cached input, $2.70 per 1M output. 5	Five times the input price of StepFun's own Step 3.7 Flash. 5
Where it stands	StepFun says 44 puts it in the top three open models, at one-eighth of Claude Opus 5's cost per task. 2	Behind GPT-6 Astra and Claude Opus 5 on the hardest coding and document tasks. 1
The date that matters	Weights on October 15. Until then the model is API-only, and Artificial Analysis lists it as proprietary. 1 3	Self-hosting, and fine-tuning downloaded weights, wait until mid-October.

What Step 5 Preview is

StepFun's documentation gives the model three reasoning-effort settings, tool calling, JSON Schema output and prompt caching, and says tools, search and code execution come from the application that integrates the model. 6 StepFun pitches it at work needing long context, repeated tool calls and steady progress toward a deliverable, with finance the domain it paid most attention to. 1 2

Where it wins, and where it stops

On StepFun's own table, Step 5 Preview at high effort scores 33.3% on Terminal-Bench v4, ahead of Kimi K3 at max effort (12.6%) and behind GLM-5.3 at max effort (41.9%), and leads both of those open models on StepCodeBench with 49.0%. 1 GPT-6 Astra at max effort stays above it on the hardest coding and document tasks: 57.9% on Terminal-Bench v4 and 31.0% on GDP.pdf, against 14.8% for Step 5 Preview. 1 StepFun's own summary concedes "a meaningful gap to the frontier remains" on the most difficult long-running tasks. 1

StepFun's chart of Artificial Analysis Intelligence Index against cost per task, plotted on a logarithmic cost axis with efficient, workhorse and high-compute zones — StepFun's launch page places Step 5 Preview above a "previous Pareto line" — the set of models nothing else beats on both intelligence and cost at once — and marks a 65% cost cut against the models to its right. The comparison set and the framing are the vendor's. 1

Artificial Analysis measured 99.8 output tokens per second and 2.96 seconds to the first answer token, both better than the median model in that price band. 3 It also generated 160M output tokens finishing the index, against a median of 92M; at a fixed output price, a wordier model costs more. 3

Before you deploy anything

Rate limits scale with the total you have topped up: a new account starts at 10 requests and 5M tokens per minute, and $15 in top-ups raises that to 1,000 requests and 20M tokens. 5

StepFun calls Step 5 Preview an open model, while Artificial Analysis lists it as proprietary because no weights can be downloaded today. 1 3 Until the weights land, the cheapest way to reach a 44 on that index is a hosted API, whose bill depends on how much the model writes.