Cheap per Token Is Not the Same as Cheap per Task
This is the most useful point here, and the one a price list most easily misleads you on. At $1.00 per million input and $2.70 per million output, it really is well under the medians for its tier ($1.88 and $10.00). But across the Artificial Analysis Intelligence Index evaluation it generated 160 million tokens, against a median of 92 million for the models evaluated — **close to twice the volume.** Reasoning tokens are billed as output, so a meaningful share of what the low unit price saves gets eaten by the extra talking. So the right unit for evaluating a model like this is "what did it cost to finish one of my actual typical tasks," not a price comparison. Fortunately that is easy to measure yourself: run three to five real tasks, record total tokens and total spend, and you will have something more accurate than any cross-vendor leaderboard. Another number from the same evaluation is solid, though: 33% on Terminal-Bench 4.0 against Kimi K3's 13%. That is the strongest external evidence for the claim that this really was tuned for long-horizon agent work, and terminal benchmarks of that kind are hard to inflate by being verbose.
The Date of October 15 Is What This Story Actually Turns On
The company says open weights land on October 15. On launch day, the Hugging Face repository stepfun-ai/Step-5-Preview-BF16 held a single .gitattributes file — no weights, no license, no model card, no config. That is not an accusation, it is precision: **"open weights" is currently a promise to be kept, not a delivered fact.** Until October 15, evaluating this as an API model is the only safe approach. Two things to check on that date, and you need both: whether the weights actually appear, and **under what license.** Those are separate questions, and this site's other story today supplies a ready-made comparison — Qwen-Image-2.1 shipped its weights as planned, but its license stepped back from Apache 2.0 to research-only, with commercial use requiring a separate negotiation. "Open weights" and "usable commercially" were never the same thing. What you read on the day is the license file, not the headline on the announcement. Two companies in the same country in the same week moving in opposite directions on openness is itself worth recording.
The Architectural Bet Is Depth, Not Width
Worth noting technically is the 92-layer narrow-deep structure and the 4.5% activation ratio — 27 billion activated per token out of 600 billion total. The stated rationale is direct: deeper stacks give longer information paths for implicit multi-hop reasoning during long prefill. Translated, the bet is on step-following-step reasoning inside long contexts rather than breadth of knowledge in a single call. Early evaluation results line up with that orientation — per some early write-ups on Zhihu, it does relatively well on scientific coding while showing gaps in factual knowledge and complex document handling. Which yields a clear selection guide: **if your workload is running one tool chain over a long stretch while repeatedly reading back its own earlier output, it is worth trying; if you need factually accurate Q&A or long-document comprehension, test it on your own material first.** Finally, that discrepancy: the official context figure is one million, while a third-party configuration circulating in developer tooling says 350,000 with a 64,000-token output cap, and the official announcement mentions no output cap at all. Anyone who genuinely needs the long context should trust neither number and fill it to the limit once themselves.
via: Artificial Analysis's Step 5 page, Pandaily, DataStudios, Orca Router's analysis of the open-weights promise