Kimi K3 behaves very differently across API, Claude Code, and its native client
APPSO · wechat · 2026-07-23
This hands-on comparison tests Kimi K3 in four setups for a screenshot-to-HTML task: direct API, Claude Code, Kimi’s native client, and Codex.
What happened
- Direct API: shortest path, but no process feedback; despite that, it delivered the first runnable page fastest.
- Claude Code: felt like a real coding agent because it could inspect files, use the terminal, and iterate, but the first pass generated code without actually writing files to disk.
- Native Kimi client: showed the model inside Moonshot’s own prompting, tools, and delivery flow.
- Codex baseline: used as a comparison point after the bridge to Kimi through CCSwitch kept failing with 502 errors.
Main takeaways
- The same model can behave very differently depending on the harness around it.
- Agent wrappers can add capabilities, but they also add new failure modes: file writes, tool calls, execution feedback, and verification can all break.
- The native client is valuable not because it always produces the prettiest output, but because it hides the operational complexity from normal users.
The article’s bigger point: “wrapper” is not just packaging. A harness can organize capability, create capability, and introduce new bugs at the same time.
More from coding & agent
- Gergely Orosz: Shipping 10x PRs With AI Agents, Sites Fill With Small Regressions — ducha_aiki · 2026-09-11
- Same Echo Maze prompt, three frontier models: all passed visually but shipped the same hidden bug — eyishazyer · 2026-09-11
- Astra storyboards plus Minimax H3 per-shot generation boost video success rates — Hailuo_AI · 2026-09-11
- Codex tip: use Sol with Astra and Luna sub-agents to save usage — pvncher · 2026-09-11
- agents-best-practices: a provider-neutral Agent Skill for designing and auditing agentic harnesses — tom_doerr · 2026-09-11
- Cognition's SWE-2 uses a KKT duality argument in RL to shift the effort Pareto curve — YouJiacheng · 2026-09-11