# StepFun Step 5 Preview 编程智能体实测

- 来源：elvis (@omarsar0)
- 发布时间：2026-09-21 21:59
- AIHOT 分数：41
- AIHOT 链接：https://aihot.news/items/cmubbr2us0tylrolnqj6g6gg2
- 原文链接：https://x.com/omarsar0/status/2102035017262149640

## AI 摘要

StepFun 新模型 Step 5 Preview 在编程智能体测试中表现亮眼，成本与能力处于帕累托前沿，可与 GLM 5.3、Kimi K3 等模型比肩且价格极具竞争力。

## 正文

StepFun’s new Step 5 Preview model is impressive!

Had a chance to test it early.

I've been testing it as a coding agent.

It’s on the Pareto frontier for cost vs. capability.

It is a very capable model, comparable to GLM 5.3, Kimi K3, and others, and it comes at a very competitive price.

It’s worth trying in your favorite coding agent.

I tested it in a minimal harness to see how it compares with GLM 5.3.

Here are my results.

Overall, it is very effective and feels like a model I could use for a whole range of coding tasks. One behavior that stood out is that it knows when to stop, which makes it very effective at long-horizon tasks compared with other models in this class.

I gave it and GLM 5.3 two real tasks in the same repo, at the same commit. First, a bug where numeric filters silently returned zero rows for decimals and negatives. Then a feature that needed new routes, permission gating, and a refactor of the background task supervisor.

Both models got both tasks right. Every held-out test passed with no regressions, and neither one weakened an existing test to get there.

Step 5 Preview finished, checked its work, and declared itself done. Both times.

GLM 5.3 wrote correct code both times and then kept going until the step limit ended the run.

Neither model got a follow-up prompt or a retry. Both had zero fix rounds, so the first delivery was the final delivery on both tasks.

On the bug, Step 5 Preview wrote the shorter patch, the same approach the Datasette maintainer used in the real commit. It also added its own tests without being asked.

Based on this, I would reach for it on unattended agent runs where a clear completion signal matters more than speed, on bug fixes in unfamiliar codebases, and on long-context work.

On long context, I gave it a separate test earlier in the week. I generated about 368K tokens of fake incident tickets and hid five clues inside them that together explain an outage. When I asked for the root cause, it found all five clues and connected them correctly in about 90 seconds.

Thanks to the StepFun team for partnering on this post.