Skip to content

opencode v2.3.2 fixes GPT-6.2 prompt caching billing

The release moves the system prompt out of the Responses instructions field so GPT-6.2 models stop charging it as uncached input on every step.

Draftship Editorial2 min read

opencode released v2.3.2 on 18 November 2026, and the main item is a fix for GPT-6.2 prompt caching. According to the release notes for v2.3.2, GPT-6.2 models do not cache a system prompt sent in the Responses instructions field. Every step of a run therefore re-billed that prompt as uncached input. opencode now sends the system prompt as a leading developer message on every Responses route, which means it is cached from the second step onward.

What the patch actually touches

The change is not limited to one provider. The notes list OpenAI through an API key and through ChatGPT sign-in, GitHub Copilot, Zen, Azure and OpenRouter as routes that receive the new treatment. The work is tracked in pull requests #64542 and #64544. Compaction, the step where a long session is condensed to free up context, now reads the same cached prefix through /responses/compact. A separate fix stops webfetch structured output from repeating the url and format it was called with (#64534).

The measurement behind the fix

The release includes one number. On GitHub Copilot with gpt-6.2-sol, uncached input per step fell from about 14,000 tokens to 3. The notes attribute the previous figure to the system prompt being charged again on each step. That is the whole of the quantitative evidence published: one model, one route, one before and after.

What this likely means for cost (inference)

This part is inference, not something the source states. Cached input is generally billed at a lower rate than uncached input, so a drop from roughly 14,000 uncached tokens per step to 3 suggests a meaningful reduction in the input side of a bill for long agent runs. The effect likely grows with the number of steps in a session, because the same prefix is reused on every turn. The source publishes no prices, no cache lifetime and no estimate of the money saved, so treat the size of the saving as unknown.

Open questions

The notes do not say whether other model families share the same behaviour in the instructions field, or whether opencode had already worked around it elsewhere. They do not describe how long the cached prefix survives, so it is unclear what happens to a run that pauses for a long time between steps. Moving the system prompt into a developer message is a change in how the request is structured, and the source does not discuss whether that affects how strongly a model follows those instructions. Nothing is said about the other routes beyond the list of names, so the Copilot figure is the only one you can check.

For people building apps with AI

If you use an AI coding agent, the practical lesson is that where a prompt sits in a request can change what you pay. A system prompt in the wrong field was being charged on every step without anyone noticing until the tokens were counted. When you build with a model API, check how your leading instructions are sent and whether they are being cached.

Draftship, the AI app builder, is not connected to this release. Its docs on using the AI chat and choosing a model cover how model selection works inside the builder.

Primary sources

  • v2.3.2 · opencode releases, Nov 18, 2026

This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.