Hands-on with StepFun Step-5-Preview: rock-solid agent loops, weak on 3D and aesthetics
karminski3 · x · 2026-09-25
- A hands-on review of StepFun's Step-5-Preview highlights output stability as its biggest strength — solid post-training means the model rarely flip-flops on engineering coding tasks.
- In a new "silicon traffic cop" agent test it caused zero accidents across runs; scores in backend agentic coding and the older courier test also showed very low variance.
- Surprise finding: strong agent-loop iteration — it can optimize right up to limit values. In one test it interpreted "no penalty if delivery is under 5 minutes late" as delivery time +5 minutes, exploiting the rule for more points.
- Weaknesses: front-end generation, 3D scenes, and aesthetics still trail this generation's SOTA, making flashy 3D demos hard.
More from coding & agent
- Runway MCP Lets Any AI Agent Generate Video Directly — tlakomy · 2026-09-25
- Hexis: open-source Postgres-native cognitive architecture giving LLMs memory and identity — GeretsOliver · 2026-09-25
- AgentArk, distilling multi-agent debate into a single LLM, accepted to NeurIPS 2026 — jindong_wang92 · 2026-09-25
- video-shotcraft author: Opus 5.5 makes a full code-generated video with music in one prompt — caolonghua · 2026-09-25
- Stability Matrix hang on Linux traced to 143 Windows fonts breaking Fontconfig — Signal_Care6558 · 2026-09-25
- Dev builds AI agent that calls 5 massage spas in parallel to compare prices and book slots — MarchCold7216 · 2026-09-25