StepFun's Step 5 Preview Packs 600B Parameters Into a $1-per-Million Coding Model

StepFun skipped a generation and went big. The Shanghai lab announced Step 5 Preview on September 20 with same-day API access, jumping straight past the Step 4.x numbering line to a 600-billion-parameter sparse mixture-of-experts model priced at $1.00 per million input tokens. The pitch is explicit: a flagship model for production-scale agent applications, with software development named as the first use case.

The architecture is deliberately narrow and deep rather than wide: 600 billion total parameters with roughly 27 billion active per token, arranged in a 92-layer narrow-deep Transformer layout that StepFun argues improves implicit multi-hop reasoning during long-context prefill. Context runs to 1 million tokens with native image and video input, and the training emphasis is on-policy long-horizon reinforcement learning — the kind of post-training that matters for agents that run tools over extended sessions rather than answer single prompts.

The price-to-intelligence pitch

On the Artificial Analysis Intelligence Index, Step 5 Preview scores 44 — level with Moonshot’s Kimi K3 Max, against 47 for GPT-5.6 Sol. That three-point gap comes with a roughly sevenfold price difference: output runs $2.70 per million tokens with a 95% cache discount, undercutting the dense flagships it sits just below. StepFun’s tagline for the launch, “Advancing the Pareto Frontier,” is a claim about the cost curve, not the benchmark curve — and it’s the same argument DeepSeek used to reshape the market two years ago.

The release is also a promise of open weights on October 15. A Hugging Face repository for the BF16 variant already exists but ships no weights yet — currently just a placeholder. If the date holds, teams running coding agents on their own infrastructure get a frontier-adjacent agentic model they can self-host within a month.

What to watch

Two caveats keep this in perspective. First, Artificial Analysis evaluated the model from StepFun’s API starting September 18, two days before the formal announcement — ordinary lag between a model going callable and a vendor publishing about it, but the public benchmark picture is still thin. Second, Step 5 arrives in a week where Chinese labs released three frontier-adjacent models within 48 hours, and the efficiency race is now the product strategy, not just a technical preference. For developers picking the engine behind a coding agent, the question isn’t whether Step 5 Preview matches Sol — it’s whether 44 points of intelligence at $1 per million tokens is the better trade than 47 at seven times the price. For a lot of agentic workloads, it is.