Playbooks · Coding tasks
OpenCode 2.3.2 stops re-billing the system prompt
Written November 18, 2026
OpenCode 2.3.2 sends the system prompt as a developer message so GPT-6.2 caches it. On Copilot, uncached input per step fell from about 14,000 tokens to 3.
What changed
OpenCode 2.3.2, released November 18, fixes how the system prompt reaches GPT-6.2 models. The prompt itself has not changed, and neither has the model. What changed is where the prompt sits in the request, and that placement decides whether the provider caches it or bills it in full on every step of a run.
Why the prompt was billed on every step
An agent run is not one request. Each step sends the conversation again, with the system prompt at the front, so a long task can send the same prompt dozens of times. Prompt caching is what normally keeps that affordable: the provider keeps the repeated opening and charges later steps the lower cached rate for it. The 2.3.2 notes state the problem plainly: GPT-6.2 models do not cache a system prompt sent in the Responses instructions field. OpenCode was sending the prompt in exactly that field, so every step re-billed it as uncached input, no matter how many steps had already sent the identical text.
What 2.3.2 sends instead
OpenCode now sends the system prompt as a leading developer message. The notes say the change applies on every Responses route, including OpenAI by API key and by ChatGPT sign-in, GitHub Copilot, Zen, Azure and OpenRouter. The sign-in method and the provider in the middle no longer decide whether the prompt caches: on each of those routes it is cached from the second step on. The first step still pays the full uncached price, because there is nothing cached yet. From the second step, the repeated prompt is read from the cache instead.
What the numbers mean
The release gives one measured example. On GitHub Copilot gpt-6.2-sol, uncached input per step dropped from about 14,000 tokens to 3. Read that as a per-step figure, not a per-run one, and the size of the fix follows. On a ten-step run at the old behaviour, the prompt alone accounts for roughly 140,000 uncached tokens, because each of the ten steps bills about 14,000 again. On the same run after updating, the first step bills about 14,000 and each later step bills about 3, for roughly 14,000 in total. The saving grows with the length of the run, which is why this fix matters most to the people who run long, multi-step tasks, and why a one-step question shows almost no difference. The figure is the release's own example on one Copilot model; other routes were changed the same way, but the notes give no second measurement.
Compaction and one smaller fix
Compaction inherits the fix. When a long session compacts through /responses/compact, it now reads the same cached prefix, so the compacted request does not go back to paying full price for the prompt either. That matters on the longest sessions, where compaction is routine rather than rare. The release carries one further fix, in webfetch: structured output no longer repeats the url and format it was called with, so the result handed back to the model is the fetched content without an echo of its own call arguments at the front.
What to do
Update to 2.3.2 if you run GPT-6.2 models through OpenCode, especially on multi-step work. No setting changes with it: the new placement is the default behaviour on the routes listed above. For multi-step GPT-6.2 runs this is the rare patch whose effect is visible in the token bill, so check uncached input per step before and after updating. If the per-step figure stays near its old level, the run is probably not on a GPT-6.2 Responses route, or it is too short for caching to show. The previous OpenCode release is covered in OpenCode 2.3.0: subagents move to the background.
Related playbooks
- Move OpenCode 1.x to 2.0 without losing config
- OpenCode 1 offers the upgrade to OpenCode 2
- OpenCode 2.2.6: scripted sessions, clean trees
- OpenCode 2.2.7: Copilot Auto and breaking changes
Sources
Back to: Coding playbooks hub · Also: OpenCode 2.3.0: subagents move to the background