Sources
StepFun ships Step 5 Preview: 600B sparse MoE, 27B active, 1M context, weights on October 15
Step 5 Preview is available today through StepFun's products and API, with full open weights promised for October 15, 2026. It scores 44 on the Artificial Analysis Intelligence Index at roughly one-eighth the per-task cost of Claude Opus 5, but the raw benchmark gaps are real: 67.7% on DeepSWE v1.1 against GPT-6 Astra's 74.1%, 33.3% on Terminal-Bench v4 against Astra's 57.9%, and 49.0% on StepCodeBench against Opus 5's 63.9%. The pitch is explicitly cost-per-task rather than frontier capability, so the question for builders is whether an eighth of the price buys back the Terminal-Bench deficit on your workload.
Source
↳ Follow the thread