Report: Step 5 Preview ties Kimi K3 at a fifth of the output price
StepFun's 600B reasoner already benchmarks at K3's level and beats it on agentic tasks — but there's no public endpoint or weights yet.
Copy markdown
A K3-class reasoner at a fifth of the output price
Step 5 Preview posts an Intelligence Index of 44 — level with Kimi K3 — while listing at $1/$2.70 per million tokens against K3's $3/$15. If it ships at those numbers, reasoning-heavy pipelines get roughly 5x cheaper on output.
Where it beats K3, and where it ties
On Terminal-Bench 4.0 it scores 33% to K3's 13% — a real gap on agentic, tool-driven work — and ties at 59% on SciCode. It's a ~600B-parameter MoE with about 27B active, a 1M-token context, and text and image input.
The catch: you can't run it yet
There's no public endpoint and no open weights; the only access is StepFun's internal test API, which Artificial Analysis is scoring under a preview tag. StepFun hasn't officially announced it — its docs still stop at Step 3.7 — so treat the price and specs as preliminary. This is a heads-up on what's coming, not a launch you can wire in today.