OpenCode fix cuts GPT-6.2 uncached input per step from ~14,000 tokens to 3

victor_explore · x · 2026-11-19

OpenCode v2.3.2 fixes a costly quirk: GPT-6.2 models don't cache a system prompt sent via the Responses instructions field, so agents re-billed it as uncached input on every step. OpenCode now sends it as a leading developer message across all Responses routes (OpenAI API/ChatGPT sign-in, GitHub Copilot, Zen, Azure, OpenRouter), making it cached from step two onward; /responses/compact reads the same cached prefix. On Copilot's gpt-6.2-sol, uncached input per step dropped from 14,000 tokens to 3. Also fixes webfetch structured output repeating url/format args. A useful pitfall to know for anyone building agents on the Responses API.

Original post →

More from coding & agent

coding & agent channel →