On the morning of September 20, StepFun released and fully opened its flagship model Step 5 Preview, announcing that model weights would be released on October 15. According to Jiemian News and National Business Daily reports, the model uses a sparse MoE architecture with 600 billion total parameters but only 27 billion activated, supports a 1-million-token context window, natively handles text and image input, and is positioned as a flagship foundation model for real-world agentic tasks.
The news is not in the parameters; it is in t
[1][2]he pricing logic. According to reports citing Cailianshe and The Paper, Step 5 Preview has entered the global open-source top three on the third-party Artificial Analysis Intelligence Index — and the more striking figure is the cost claim: per reports citing the STAR Market Daily, the model's per-task cost is about one-eighth that of Claude Opus 5. Of 600 billion parameters, only 2.7 billion are activated per step; that sparsity ratio is the entire cost story. Treating "how much you activate" as a competitive dimension is the open-source camp's direct answer to closed frontier labs.
The timing is worth recording. On the same days that Google was reported to be testing Gemini 4 Pro under an anonymous alias and OpenAI and Anthropic were adjusting their release cadences, a Chinese company put "API fully open + open-source date announced" in the same launch. Closed labs compete on secrecy; the open-source side competes on openness — Step 5 Preview puts the tension on the table: a flagship that "leaks" versus a flagship that pre-announces its open-sourcing.
The attribution boundaries need stating. The AA index is a composite from the third-party evaluator Artificial Analysis; "top three open-source" is that ranking's framing, not an official rank. The "one-eighth of Claude Opus 5 per-task cost" comes from news reports citing the STAR Market Daily, not from an official price sheet — StepFun officially confirmed the architecture, context window and release dates, nothing more. Both are verifiable third-party statements; neither should be read as official data.
For developers, two things are actually testable: whether the 1-million-token context window is genuinely usable in the API, and whether the weights arrive on October 15. Sparse-MoE inference efficiency and agentic tool-calling behavior need community re-testing once the weights ship — that is the discipline of open models, and it is more convincing than the launch event itself.
As of writing, the API and Studio access are open. Whatever happens to its leaderboard position next week, this launch has already done one thing: it moved the Chinese open-source price war from "price per million tokens" to "per-task cost" — a unit of measurement aimed directly at closed flagship models.
None of this is a claim that Step 5 Preview beats its closed rivals on raw capability; open-weight models rarely lead leaderboards for long, and the AA composite favors cost-normalized performance by design. The point is structural. Extreme sparse MoE — 27 billion of 600 billion parameters activated — is where the open-source economics live: you pay for what runs, not what exists. Whether that trade-off holds on real agentic workloads, with tool calls and long contexts, is exactly what the October 15 release will let the community test.
[1][2]