StepFun releases Step 5 Preview: 600B MoE with 1M context at $1/$2.70 per 1M tokens

StepFun has released Step 5 Preview, a sparse Mixture-of-Experts model with 600B total parameters and 27B active per token, a 1.0M-token context window, and API pricing of $1.00 per 1M input tokens and $2.70 per 1M output tokens. The OpenRouter listing dates the release to October 8, 2026.

Key specifications

Spec	Step 5 Preview
Architecture	Sparse Mixture-of-Experts
Total / active parameters	600B / 27B
Context window	1.0M tokens
Input price	$1.00 per 1M tokens
Output price	$2.70 per 1M tokens
Cache read price	$0.05 per 1M tokens
Release date (per listing)	Oct 8, 2026

Benchmark scores, training cutoff date, and a technical report were not included in the listing. Input and output modalities are not specified in the available page text.

Positioning

StepFun describes Step 5 Preview as its flagship model for agentic work. According to the company's description, it performs strongly in software engineering and professional knowledge work, with particular strength in finance. StepFun says it is designed for extended tasks spanning large codebases and documents, using tools and refining results over multiple steps. These are company claims, and no independent evaluations have been published.

Serving performance

OpenRouter's metrics, measured over the last three days, show StepFun as the only listed provider:

• Latency: 1.68s P50
• Throughput: 137 tokens per second P50
• Uptime: 100.00%, meaning requests were routed to a provider
• Availability: 94.41%, which counts errors and empty responses against the model

The gap between uptime and availability suggests some failed or empty responses. That is plausible for a preview-stage launch, but the data covers only a short window.

Placement in StepFun's lineup

Step 5 Preview sits well above StepFun's other listed models on OpenRouter:

• Step 3.7 Flash: a multimodal MoE with a 196B-parameter language backbone, roughly 11B active, and a 256K context window (listed as 262K). It costs $0.20 input and $1.15 output per 1M tokens.
• Step 3.5 Flash: a 196B-parameter reasoning model with 11B active, a 262K context window, and pricing of $0.10 input and $0.30 output per 1M tokens. StepFun calls it its most capable open-source foundation model.

Versus Step 3.7 Flash, Step 5 Preview has about 3x the total parameters, about 2.5x the active parameters, and about 3.8x the context length. Input pricing is 5x higher and output pricing about 2.3x higher. Versus Step 3.5 Flash, input is 10x and output 9x higher.

The listing does not say whether Step 5 Preview's weights will be released, unlike Step 3.5 Flash, which StepFun describes as open source.

What this means

Step 5 Preview moves StepFun from efficiency-focused models toward a larger agentic flagship. The 27B active parameter count keeps per-token compute moderate for a 600B model, and the $2.70 output price is low for a 1M-context agentic model. Cache reads at $0.05 per 1M tokens matter for the workloads StepFun targets, since repeated passes over large codebases and documents are mostly cached input.

The unknowns limit any judgment for now. Without published benchmarks, there is no way to verify the claimed strengths in software engineering and finance against other agentic models. The "Preview" label and the 94.41% availability figure also suggest the model is not yet at production stability. Teams evaluating it should test long-context retrieval and tool-use reliability themselves before moving workloads off a cheaper Flash-tier model.