Opus 4.6 reasoning regression: fails simple file access tasks that Haiku handles correctly

Summary

Opus 4.6 (1M context) is exhibiting significant reasoning regression — specifically, failing on simple tasks where smaller models (Haiku, Sonnet) do not. This is not a capability ceiling issue; it is a regression from expected Opus-tier behavior.

Observed behavior

When a file exists at a known absolute path and the model has full filesystem access, Opus 4.6 diagnosed the situation as a "tooling limitation" rather than simply reading the file. It produced multi-step diagnostic frameworks, offered "fix options," and explained why the file was inaccessible — without ever attempting to read it.

The file was accessible. The model had the Read tool. No actual barrier existed.

A simpler model would have read the file.

Impact

• Opus produces elaborate explanations for problems that don't exist
• It mistakes its own inaction for environmental constraints
• This compounds in longer sessions: wrong diagnosis → wrong fix → wasted turns

Expected behavior

When a tool is available and a path is known, use the tool. Do not theorize about access barriers before attempting access.

Severity

High. If Opus is less reliable than Haiku on straightforward tool-use tasks, the model tier premium is unjustified and the regression is user-visible.

Steps to reproduce

1. Start a Claude Code session in directory A
2. Tell the model a relevant file exists at directory B (different from cwd)
3. Ask the model to use context from that file
4. Observe whether the model reads the file or theorizes about why it cannot