A Chinese Lab's New Model Just Jumped 24 Spots on OpenRouter's Usage Chart. Its Open Weights Aren't Out Yet.

StepFun's Step-5 Preview moved from 632.2 billion to 3.22 trillion tokens processed on OpenRouter this week, a 409.3 percent jump that lifted it 24 places on the aggregator's model-usage leaderboard to the #11 spot overall, with 2.1 percent of total platform volume. That is the single largest usage swing of any model tracked in this period, ahead of established names like Anthropic's Claude Haiku 5.5 (+72.8 percent) and Meta's Muse Spark 1.3 (+51.7 percent). It is a striking debut for a model from a Chinese lab most Western developers have not heard of — but the part of the story that will actually determine whether it matters is still five days out.

A 600-billion-parameter model that undercuts GPT-5.6 Sol

StepFun opened API access to Step-5 Preview on September 20, 2026, describing it as a sparse mixture-of-experts model with 600 billion total parameters and 27 billion active per token — roughly 4.5 percent of the network engaged on any given pass — plus a 1-million-token context window and text-and-image input. On Artificial Analysis's Intelligence Index, it scores 44, putting it in a tie with Kimi K3, Grok 4.6, and GPT-5.6 Sol at its standard reasoning effort (Sol needs its highest, "xhigh," effort setting to also reach 44). The gap shows up in price, not capability: Step-5 Preview charges $1 per million input tokens and $2.70 per million output tokens, against roughly $4 input and $20 output for Sol at its current promotional pricing — Artificial Analysis put Step-5's cost per completed Intelligence Index task at $0.72, versus $1.18 for Sol, and as of its September 21 analysis, no other model scoring 44 or higher ran cheaper per task. StepFun also offers up to a 95 percent prompt-cache discount, which can push effective input cost toward roughly $0.05 per million tokens for cache-heavy workloads like long-horizon coding agents, the use case the company is explicitly targeting the model at.

The open-weight promise is still a promise

StepFun has said it will release Step-5 Preview's weights on October 15, 2026 — five days from this writing. Right now, that promise is unverifiable in the most literal sense: the Hugging Face repository StepFun has designated for the release, `stepfun-ai/Step-5-Preview-BF16`, exists but contains nothing but a `.gitattributes` file — no weights, no model card, no configuration, and no license terms published anywhere yet. Everything currently driving Step-5 Preview's OpenRouter surge — the pricing, the benchmark score, the 600B/27B architecture claim itself — is running through StepFun's hosted API, not an open release, and all of it remains a vendor claim that cannot be independently checked against the weights until they actually land. The gap between "promised open-weight model" and "open-weight model" is exactly where several other 2026 releases have stumbled or been delayed, and nothing about this week's usage spike tells you which way October 15 will go.

A crowded field of cheap, non-Western frontier contenders

Step-5 Preview's jump is also a data point in a broader pattern already visible in this week's OpenRouter numbers: of the platform's top 10 models by raw token volume, seven are from Chinese labs — DeepSeek's V4.1 Flash alone holds 24.8 percent of all platform tokens, with Z.ai's GLM-5.3 Flash, Xiaomi's MiMo V2.6 Flash, and Tencent's HY4 Preview all in the top four behind it. OpenAI's GPT-6 Luna sits fifth at 4.6 percent share, and Anthropic's Claude Opus 5.5 is eighth at 2.8 percent — meaning Step-5 Preview's new #11 position already puts it ahead of several models from labs with far larger public profiles. None of this is new as a trend; what's new is a lab without an existing OpenRouter foothold breaking into the top 15 in a single week, on a model that is not yet available to download, fine-tune, or audit independently.

The caution here cuts against treating this week's spike as confirmation of anything beyond usage. Step-5 Preview's benchmark comparisons so far come from Artificial Analysis's own test harness and StepFun's own published pricing — real, specific numbers, but not yet cross-checked by an open weight release the way, for instance, a Hugging Face community distill or quantization might confirm a model's claimed architecture within days of launch, as happened with Google's EmbeddingGemma 2 last week. Step-5 Preview's actual test of that kind arrives on October 15, if the date holds; until then, this is a lead on a usage chart, not yet a story about an open model.