News · 2026-10-09

StepFun’s Step 5 Preview reaches OpenRouter with a million-token context

OpenRouter lists StepFun’s Step 5 Preview as a hosted model with a one-million-token context window and posted rates of $1 per million input tokens and $2.70 per million output tokens. The listing gives October 8 as its release date, while Artificial Analysis records an earlier September appearance. The development is verified hosted availability; the dossier does not settle official downloadable-weight status.

Key facts

• OpenRouter lists 600 billion total parameters, with 27 billion active per token.
• The platform dates its listing to October 8, 2026; Artificial Analysis records September 18 for the model.
• StepFun’s model accepts text, images, and video and produces text, according to the listing.
• The primary source for hosted availability is OpenRouter’s Step 5 Preview page.

OpenRouter describes Step 5 as StepFun’s “flagship model for agentic work.” The provider page targets extended tasks across codebases and documents, with tools and repeated refinement. That is a published positioning statement, rather than a guarantee of successful autonomous work. Readers are choosing a hosted endpoint with testable specifications, while a settled local deployment recipe remains unverified.

The model uses a mixture of experts. Instead of applying every part of the network to every token, a router selects a subset of specialized components. An analogy is a large consulting firm that sends a small team to each assignment. The team working on one task is smaller than the whole company, but the full company still exists and requires infrastructure.

That distinction prevents a common misunderstanding. Twenty-seven billion active parameters does not mean that Step 5 is stored or hosted like an ordinary twenty-seven-billion-parameter model. It describes which computation is active for a token. The total parameter count remains a separate quantity, and neither count establishes consumer-hardware requirements.

The context window is another separate quantity. One million tokens means a request can accommodate a large amount of text or other encoded input within the service’s stated limit. It is not proof that the model retrieves every fact reliably across that entire span. OpenRouter also lists up to 64,000 completion tokens. A large output allowance can support extended reasoning or lengthy deliverables, but it can increase waiting time and expense.

The checked prices include a cache-read rate of five cents per million tokens. StepFun’s own China-platform pricing documentation provides a separate local-currency schedule. These are posted service rates, not an independently measured price for finishing a particular coding project. Builders should compare complete tasks under comparable instructions, rather than input rates alone.

Artificial Analysis supplies a useful cost caveat. Its model record reports an Intelligence Index score of 44 against a comparison-group median of 26. More revealing for operating cost, it reports 160 million output tokens during evaluation, compared with an 81 million median. That is roughly twice the comparison group’s median output volume. The result belongs to that evaluation and does not prove every user will see the same expansion.

A lower output-token price can be offset by longer answers or more reasoning. Consider a contractor charging less per hour while taking more hours to complete the job. Neither the hourly rate nor the hours alone tells you the invoice. The same logic motivates comparing accuracy, task completion, latency, and total spend together, as explained in inference economics.

Artificial Analysis dates the model to September 18, whereas OpenRouter says October 8. The evidence therefore supports a new listing on this distribution platform, rather than an unqualified claim that the model first appeared in public that day. Its reported evaluation score is third-party evidence, while the architecture and interface specifications come from the service listing.

The weight situation is unresolved. The dossier found disagreement between an evaluator’s proprietary classification and the model organization’s visible listing, then a community link to an apparent official weight repository that was not inspected directly. It would be premature to call the model definitively closed or definitively downloadable. No verified disk-size or local running-memory figure is supplied, and neither is guessed here.

The Hacker News discussion includes interest in coding-agent use and skepticism about local deployment of the full architecture. Those are reasonable questions, not benchmark results. The concrete decision today is whether to test the hosted endpoint against a real workflow. For that choice, long-context behavior, output volume, and successful task completion matter more than the appeal of a sparse model’s active-parameter headline.

Primary source, verified: read the paper →

Ground Truth. (2026, October 9). StepFun’s Step 5 Preview reaches OpenRouter with a million-token context. Ground Truth. https://groundtruth.day/news/stepfun-step-5-preview-openrouter-hosted-access.html

@misc{groundtruth:stepfun-step-5-preview-openrouter-hosted-access,
  title  = {StepFun’s Step 5 Preview reaches OpenRouter with a million-token context},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {oct},
  url    = {https://groundtruth.day/news/stepfun-step-5-preview-openrouter-hosted-access.html}
}

Topics: models · stepfun · openrouter · agents · mixture-of-experts