OpenCode shipped version 2.3.2 on November 18, 2026, and the change is a single line of request shaping: the GPT-6.2 system prompt now travels as a leading developer message instead of sitting in the Responses instructions field. That one move makes the biggest repeated block in every agent session cacheable from the second step, turning a bill that re-charged the full prompt on every turn into one that charges it once.
The cause was request shape, not model behaviour under load. GPT-6.2 does not cache a system prompt placed in the Responses instructions field, and OpenCode's built-in openai provider talks to that interface. So each step re-billed the entire prefix as uncached input, quietly, at the most expensive rate on the invoice.
Why the Prefix Was the Whole Bill
Agent prefixes are not small. Every OpenCode agent ships its own system prompt plus its tool definitions, and a real session can run dozens of steps. A 20,000-token prefix across 50 steps is 1,000,000 input tokens, and at the gpt-6.2-sol rate that is about $2.00 for a conversation in which nothing structural changed between turns.
- gpt-6.2-sol: $2.00 input, $10.00 output, $0.10 cached input per 1M tokens
- gpt-6.2-astra: $8.00 input, $40.00 output, $0.40 cached input per 1M tokens
- Cache writes: $2.50 per 1M for gpt-6.2-sol, 30-minute retention, billed separately
That arithmetic is why prompt caching has become the main cost lever for terminal agents. The model is charged for what it reads, and an agent re-reads its own scaffolding on every turn whether or not the user typed anything new.
Cached from step two, the same 50-step run lands near $0.04 for the prefix plus $0.098 of uncached remainder, roughly $0.14 before cache-write fees. Same work, same model, a difference of about 14x.
Two Releases in One Day
Version 2.3.2 was not alone on the shelf. Version 2.3.1 landed the same day and changes what happens when a command outlives its timeout, which matters more than it sounds for agents that start builds and test suites.
- Timed-out shell commands move to the background instead of being killed
- timeout: 0 now means the 120-second default rather than no limit
- Compaction summaries are written in the background, disable with compaction.background: false
- GPT 6.2 Sol and GPT 6.2 Astra appear in the model picker after ChatGPT sign-in
The caching fix covers the OpenAI API-key and ChatGPT sign-in routes, GitHub Copilot, Zen, Azure, and OpenRouter. It does not name custom baseURL gateways, so anyone pointing OpenCode at a proxy has to verify the behaviour instead of assuming it.
Verify, Do Not Trust the Changelog
Three checks settle it. Confirm opencode --version reports 2.3.2 or newer, confirm requests actually travel the Responses route and are not being handled by an older provider plugin, then run a few steps and read the usage log. Cached input should appear from the second step onward, and cache writes should appear alongside it.
Set expectations by session shape. A short chat saves almost nothing, because the repeated prefix only gets expensive once the system prompt and tool definitions are long and the run has many steps. The teams that feel this fix are the ones running hundreds of agent turns a day on a fat prompt.
Read as a feature, 2.3.2 is minor. Read as a billing correction, it is the most consequential line in the release notes, and the number worth measuring afterward is not latency but the cached-input line in next month's invoice.
Comments