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.
More from coding & agent
- Cognition's Poke experiment: pitch your idea by text, win up to $10,000 in Devin credits — marvinvonhagen · 2026-11-19
- Google Antigravity ships huge week: parallel multi-agent runs, SSH dev, PR cards — clmt · 2026-11-19
- GPT-6.2 randomly abandons tasks mid-step, users resort to self-set timers to keep it going — bofors14 · 2026-11-19
- Remotion ships real motion blur, addable via AI agents with one prompt — Vjeux · 2026-11-19
- SREGym open-sources a benchmark replaying real incidents like Cloudflare and Reddit outages for SRE agents — tianyin_xu · 2026-11-19
- An LLM is not an agent: the 10 building blocks of agentic systems, from orchestration to RAG — goyalshaliniuk · 2026-11-19