What models do you use?

Vesper Code uses different models for different tasks. The orchestrator coordinates; subagents handle specific jobs.

The main agent ("Vesper") coordinates everything:

• ChatGPT Default — GPT-5.3 Codex
• ChatGPT Max — GPT-5.5 for thinking-heavy work
• Kimi — Kimi for Coding

The orchestrator spawns these for specific jobs:

• Code editing — the active provider's main or thinking model
• Thinking / reasoning — GPT-5.5 in ChatGPT Max mode, or Kimi for Coding in Kimi mode
• Code review — the active provider's review model
• File discovery — the active provider's fast model
• Terminal commands — the active provider's fast model
• Web / docs research — the active provider's fast model

Max mode runs multiple implementations in parallel and picks the best one. Default mode runs a single implementation pass.

File rewrites use speculative decoding from Relace AI.