OpenCode v2.3.2 Fixes GPT-6.2 System Prompt Caching Issue
github-actions[bot] · ghdev · 2026-11-18
OpenCode has released v2.3.2 with a key fix: GPT-6.2 does not cache system prompts sent via the Responses API instructions field, causing every step to be billed repeatedly as uncached input. OpenCode now sends the system prompt as a leading developer message across all Responses routes (OpenAI API key and ChatGPT login, GitHub Copilot, Zen, Azure, OpenRouter), so caching kicks in from the second step onward. On GitHub Copilot's gpt-6.2-sol, uncached input per step drops from roughly 14,000 tokens to 3. Also fixes a minor webfetch issue where url/format parameters were repeatedly echoed in structured output.
More from coding & agent
- Team's hands-on with MCP Apps and Events: widgets work well, event push lags behind — lemuridaelabs · 2026-11-18
- Anatomy: open-source agent skill for isometric machine illustrations hits 27.9k GitHub stars — tom_doerr · 2026-11-18
- Dev builds real-time underwater survival game in 4 days, all code written by AI — Angaisb_ · 2026-11-18
- Shopify Fall '26 Editions goes agentic: Shop Pay checkout lands in ChatGPT, Claude and more — jh3yy · 2026-11-18
- Cognition launches Devin Bots: one bot leads a team of agents, merging thousands of PRs a week — floguo · 2026-11-18
- StarSling bets $1,000 it can cut your CI bill by 10%+, cites 10x speedup benchmarks — Rasmic · 2026-11-18