AitopkitChinese AI lab Stepfun has open-sourced Step 5 Preview on October 15, 2026 — a 600-billion-parameter sparse mixture-of-experts model that activates only 27 billion parameters per token, with a 1-million-token context window and native text-and-vision input. Both BF16 and FP8 weights were released on Hugging Face and ModelScope, alongside a technical blog post.
Real-world traction is hard to ignore: as of October 15, Step 5 Preview sits at #2 on OpenRouter's weekly and daily token-usage charts, with roughly 18.8 trillion tokens processed in seven days. On Artificial Analysis's Intelligence Index it scored 44 points, placing it among the top three open-source models globally.
The cost story is the hook: Step 5 Preview averages $0.71 per task in Artificial Analysis's evaluation data — about one-eighth of Claude Opus 5's $5.86 — a striking cost-to-performance ratio for developers running agentic workloads at scale.
The usual caveat applies: leaderboard scores and per-task costs reflect the benchmark's methodology, not your workload. Verify accuracy, latency, and real API pricing on your own tasks before sending it production traffic.
With Meta's Muse Spark Open and NVIDIA's Nemotron 3.5 dropping the same week, the open-weight race is no longer a slow burn — it's the main event.
The MoE architecture is doing quiet heavy lifting here. Activating only 27 billion of 600 billion parameters per token is the trick that makes a model this large affordable to run — you get the capacity of a giant with the inference bill of a much smaller model. Combined with a 1-million-token context window, Step 5 Preview is clearly aimed at agentic workloads: long-running tasks where the model needs to hold enormous context and act over many steps. That is exactly where the $0.71-per-task figure bites — agents multiply token usage, so per-task economics decide which models get deployed at scale.
The China dimension is impossible to ignore. Stepfun, DeepSeek, and a wave of Chinese labs are now driving the open-weight frontier — releasing top-tier models freely while US labs increasingly keep their best behind APIs. It is a fascinating inversion: the open ecosystem's cutting edge is being set in Hangzhou and Shanghai as much as San Francisco. For developers, the practical effect is wonderful — genuine choice, falling prices, and models you can actually inspect and run yourself.
The strategic question is what "open" buys the releasers. Open weights do not directly make money; they build ecosystems, attract talent, and set standards that the releaser's commercial offerings then monetize. Meta proved the playbook with Llama. Stepfun's bet is that being the cost-performance leader in open weights earns it a permanent seat at the table — and the #2 spot on OpenRouter's usage charts suggests developers are already voting with their tokens.
The deeper shift is who gets to play. When frontier-class open models cost an eighth of proprietary ones per task, the moat moves from model access to everything around it — data, distribution, and workflow integration. The labs know this; it is why the open-weight race is accelerating even as it cannibalizes API revenue.