StepFun unveils Step 5 Preview, a 600B-parameter MoE model with a 1M-token context window

StepFun has launched Step 5 Preview, a flagship model built for long-horizon agentic work, software engineering, professional knowledge tasks and finance. The central pitch is cost: the company says the model delivers intelligence on par with rivals at a substantially lower per-task cost, what it calls a “Pareto frontier.” The model is available today as a hosted API and on StepFun’s platform; self-hosting will have to wait for the open-weights release, slated for 15 October 2026. A back-of-the-envelope calculation puts the 600 billion parameters at roughly 1.2 TB in BF16 before KV cache, so multi-GPU hardware is a prerequisite.

Narrow and deep architecture

Step 5 Preview is a sparse mixture-of-experts model with about 600 billion total parameters that activates roughly 27 billion per token — about 4.5% of the weights. Bucking the usual scaling playbook, StepFun chose a narrow, deep layout of 92 transformer layers. The research team says this creates longer paths for multi-hop latent reasoning, which matters during long prefill when agents are searching, executing code and reading tool returns. Training leans on on-policy reinforcement learning at long horizon, with bit-level alignment between training and inference in the MoE routing, plus MTP-3 speculative decoding, FP8 MoE and KV-cache dropping. The company reports more than a 3× end-to-end speedup in long-horizon RL.

Performance: self-reported versus independent

In coding benchmarks StepFun reports 67.7 on DeepSWE v1.1, 49.0 on its internal StepCodeBench and 80.5 on ProgramBench; GPT-6 Astra and Claude Opus 5 lead on all three. The model ran at High effort while the competitors ran at Max. In two 24-hour agent experiments, the model tuned an H100 kernel to 508 TFLOPS versus 493 for Claude Opus 5, and lifted Qwen3-30B-A3B on AIME24 from 53.3% to 60% through automated post-training. Artificial Analysis’s independent evaluation gave a score of 44 on its intelligence index, where the median for similarly priced reasoning models sits at 24; output throughput measured 99.8 tokens per second on StepFun’s API. There is a caveat: the benchmark run produced 160 million output tokens against a median of 92 million, and the verbose reasoning eats into some of the per-token savings.

Pricing and context window

The API is priced at $1.00 per million input tokens and $2.70 per million output tokens. Artificial Analysis notes medians of $1.88 for input and $10.00 for output among comparable models, so the discount is clear on input and especially on output. The context window spans 1 million tokens with support for text, image and video input and text-only output; low/medium/high thinking modes, streaming, tool calling, JSON Mode, JSON Schema and prompt caching are available. Official documentation also mentions integration with Claude Code via Step Plan and the ability to coordinate 950 web fetches in a single agent action.

Open weights in October

Open weights are expected to land on 15 October 2026, at which point self-hosting on suitable hardware becomes possible. Until then the model is accessible only as a hosted service. StepFun has not published additional independent benchmarks beyond those cited, and the internal StepCodeBench has not undergone external validation; accordingly, the claimed gap against GPT-6 Astra and Claude Opus 5 remains a manufacturer claim.