Description

Hi Team,

I am not sure whether this is an Anthropic-side issue or something specific to OpenCode, but I am seeing unexpectedly high token consumption with Opus 4.5.

In practice, the context window is jumping to ~100k tokens within 5-10 prompts, with compactions occurring very frequently. Context appears to grow by roughly 15–20k tokens per prompt. This is using a Claude Premium Business account.

For comparison, I have parallel sessions running on the same codebase and modules using OpenAI via OpenCode (OpenAI Pro subscription with the OAuth plugin). Those sessions typically run for 2–2.5 hours without a single compaction.

The agent setup, repository, and active modules are the same, so this does not appear to be driven by differences in code size or scope. Historically, this workflow allowed for longer chat → plan → implement cycles in a single session without aggressive context growth.

I wanted to flag this in case there has been a recent change in OpenCode behavior, agent configuration, or default context handling that might explain the difference.

Kind regards,
Tarek

OpenCode version

1.0.223

Steps to reproduce

Nothing specific just generate chat, plan and execute on codebase.

Screenshot and/or share link

No response

Operating System

macOS 26.2

Terminal

vscode