StepFun launches Step5 Preview: 600B MoE with 27B active params, claims 1/8 the task cost of Claude Opus 5
智东西 · wechat · 2026-09-20
StepFun released Step5 Preview, a 600B-parameter MoE flagship (27B active) with 1M-token context and native vision input, fully available now and open-sourcing Oct 15. It scores 44 on Artificial Analysis (3rd among open models) at $0.71 per task — 1/8 the cost of Claude Opus 5.
- Hands-on: stepfun's benchmarks show it beating Kimi K3 and GLM-5.3 on long-horizon agent CLI, FrontierFinance, and DRACO evals, behind only GPT-6 Astra and Claude Opus 5; 49 on the internal StepCodeBench (553 repos).
- Official demos: optimized an MLA GPU kernel to 508 TFLOPS on one H100 in 22 hours; post-trained Qwen3-30B-A3B from 53.3% to 60% on AIME24 in 24 hours; 3000+ turns in Pokémon Red; autonomously rebuilt an ESP32-S3 dev board.
- Positioning: StepFun argues the next scaling race is about compute-conversion efficiency, not raw compute.
Related event: StepFun Launches Step 5 Preview: 600B MoE Flagship, Open-Weights on Oct 15(17 posts)→
More from Models
- The Zvi opens reaction thread on Gemini 4 Carbon: is it worth extended coverage? — TheZvi · 2026-11-12
- Walden Robotics' Sage model quadruples robot uptime on Toyota factory lines — adnothing · 2026-11-12
- Gemini won't let you keep chat history while opting out of AI training — Google bundles both into one setting — arshdeep · 2026-11-12
- Gemini 4 Carbon Flex tier now costs the same per token as GPT 6.1 Terra — mariofilhoml · 2026-11-12
- Real-world test: Kimi K3.1 Pro costs 5.4x less than Opus 5.5 per task despite 2x list-price gap — Arindam_1729 · 2026-11-12
- Claude Fable 5.2 tops Agent Arena at +30.4%, 1.5x the gain of runner-up Opus 5.5 — daniel_mac8 · 2026-11-12