[BUG] [MODEL] claude-opus-4-5-20251101 - 77K apparent working context??

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

The working context with Opus 4.5 allows me 77K ONLY. The Autocompact buffer appears to automatically reserve 77.0k tokens (38.5%), but with Free space: 46k (23.0%), the autocompact is being triggered - this gives me a total working context of 77K for a model which is supposed to be 200K. I appreciate that the autocompact buffer requires some reserved space, but to not have access to 61.5% of my context allowance is entirely unworkable? The below is the /context immediately prior to autocompacting.
❯ /context
⎿ Context Usage
⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛀ ⛁ claude-opus-4-5-20251101 · 77k/200k tokens (39%)
⛁ ⛁ ⛁ ⛀ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ System prompt: 2.8k tokens (1.4%)
⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ System tools: 15.4k tokens (7.7%)
⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛶ ⛁ Memory files: 8.5k tokens (4.2%)
⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛁ Skills: 398 tokens (0.2%)
⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛁ Messages: 49.9k tokens (25.0%)
⛶ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛶ Free space: 46k (23.0%)
⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ Autocompact buffer: 77.0k tokens (38.5%)
⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝
⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝

What Should Happen?

I would expect to have access to a significantly larger working context allowance when paying for the use of a supposedly 200K frontier LLM model? The 77K for the autocompact beffer appears to be the same as the buffer for the Sonnet 4.5 with 1M context model, where it is only 7.7% of the working context so obviously much more manageable? Having exactly the same sized autocompact buffer with a model using a context window which is only 20% the size of the 1M context model seems mad - why doesn't the required autocompact buffer scale? And why is it triggering when I apparently also have Free space of: 46k (23.0%)?

Error Messages/Logs

Steps to Reproduce

The steps to reproduce just involve using 77K of the context window/allowance? I have a Premium Teams seat, in case that has any relevance here?

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.5 (Claude Code)

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

WSL (Windows Subsystem for Linux)

Additional Information

I seems to be a fundamental issue with Claude Code and the autocompact buffer? Not sure what else to add?