The user tested the Step 5 Preview model and found it performs well on long-sequence tasks in frontend coding, comparing it favorably with models like Fable and GPT5-Sol.
Step This model's progress this time is pretty significant. I tested the Step 5 Preview model, and the overall effect feels quite good. The testing mainly involved having it design the MCP mode for OpenCLI, while also comparing it with Fable + GPT5.6 Sol. Additionally, with oneshot, it implemented a side-scrolling shooter game, completing the entire process in one seamless go. The performance on long-sequence tasks is solid, and the frontend page has a pretty nice pixel art style. Overall, it performs well on long-sequence tasks in frontend coding. A comparison of the three models' basic MCP refactoring design proposals for OpenCLI: On the big picture, everyone basically stayed consistent, all agreeing to retain the original core, wrap it with a lightweight shell, and support both CLI and MCP. Breaking it down: Fable: Fable is indeed impressive; it provided a complete engineering justification. The evidence is the most dense, the plan the most specific, and the roadmap the most verifiable. Another standout aspect is its proactive deep dive into more advanced directions, boldly proposing forward-looking practices in the entire engineering evolution. It even did something few models do, akin to critical thinking: proactively refuting its own task. Water clearly delineates the layers, separating "existing assets in CLI form" and "pure new capabilities brought by MCP" into two distinct parts. But the overall design lacks forward-thinking. Additionally, it proactively broke down OpenCLI's underlying session -> tab -> target runtime dynamic tree structure, and how it maps to the new MCP form. GPT5-Sol: Nails the engineering boundaries extremely well, with very comprehensive constraint conditions. It proactively analyzed potential issues in async and concurrent environments. However, the feature analysis differentiating CLI from MCP is insufficient, as MCP can introduce more advanced functions and features. Also, the implementation path for evolution isn't very clear. Gave them scores: Fable is indeed strong, but for a domestic model to achieve this effect is actually pretty good: Executable Engineering Decision Document: Fable (9) > Sol (7) = Water (7) Insight and Ceiling: Fable (9) > Water (7) = GPT5-Sol (7)
StepFun introduces Step 5 Preview, a new flagship AI model for agentic work with frontier-level performance in software engineering and finance, featuring a 600B parameter scale.
Matt Shumer reports early access to GPT-5.6 Sol, noting it is impressive but that Fable outperforms it on most tasks and is more agentic, with a full review promised on launch day.
StepFun AI has released a preview version of their Step-5 model in BF16 format on Hugging Face, indicating a promising development in AI model accessibility and performance.