StepFun Ships Step 5 Preview, a 600B Agent Model

Chinese lab StepFun released Step 5 Preview on 20 September, a sparse mixture-of-experts model with about 600 billion total parameters and roughly 27 billion active per token — around 4.5% of the weights on any given forward pass. The hosted API went live the same day, and the company says it will open the weights on 15 October. The model is aimed squarely at long-horizon agent work rather than chat.

The architecture leans narrow and deep: 92 transformer layers, which StepFun argues gives longer information paths for the implicit multi-hop reasoning that agents need when they are searching, running code and folding tool output back into context. The one-million-token context window is served by what the lab calls Sparse GQA, a block-wise token-merging scheme that uses sparse indexing to admit only the history relevant to the current step into the attention computation. Inputs can be text, images or video; output is text. The API exposes three reasoning effort levels, streaming, tool calling, JSON mode and schema, and prompt caching.

Training was on-policy long-horizon reinforcement learning, which is expensive to run at this scale. StepFun credits a stack of systems work for making it tractable — bit-wise train-inference alignment for MoE routing, MTP-3 speculative decoding, FP8 MoE and KV-cache offloading — and reports more than a 3x end-to-end speedup on long-horizon RL as a result.

On the benchmarks StepFun chose to lead with, Step 5 Preview lands between the two Western flagships. It scores 66.4 on FrontierFinance against 69.7 for Claude Opus 5 and 55.0 for GPT-6 Astra, and 83.3 on DRACO against 87.6 and 76.8. Two caveats travel with those numbers: they are vendor-reported, and they run Step 5 at High effort while the comparison models run at Max. Independent evaluation is thinner — Artificial Analysis currently places the model at 44 on its Intelligence Index, above the median for reasoning models in its tier but well short of the frontier.

Coding is where the gap is plainest, and StepFun published those figures without rival columns.

Pricing is the loudest part of the release: $1.00 per million input tokens on a cache miss, $0.05 on a cache hit, and $2.70 per million output tokens, against a comparable-tier average the company puts at $1.88 input and $10.00 output. The list price is not the bill, though. Reasoning models are judged on what it costs to finish a task, not what a token costs, and StepFun's own materials concede that verbose reasoning generation eats into the per-token saving. Until Artificial Analysis publishes a cost-to-run figure for Step 5, the honest read is that it is cheap per token and unproven per task.

Self-hosting has to wait for the 15 October weight drop, and it will not be casual: the model runs to roughly 1.2TB in BF16 before any KV cache. StepFun is pitching it at software engineering, professional knowledge work and finance, and points to agentic runs as the proof — a single action that coordinated 950 web requests, and 24-hour experiments that included H100 kernel tuning and automated post-training optimisation. Whether that holds up outside the lab's own harness is the question the October release will answer.