Opus claude-opus-4-8 emits unparseable tool calls at ~1.5% rate; opus-4-7 and sonnet-4-6 have 0% failure rate

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?

Since claude-opus-4-8 was released, turns frequently fail with:

The model's tool call could not be parsed (retry also failed).

Analysis of local session logs (~2 weeks) shows this is specific to opus-4-8:

| Model | Assistant turns | Parse failures | Rate |
|---|---|---|---|
| claude-opus-4-8 | 9,805 | 148 | 15.1 / 1000 (~1.5%) |
| claude-sonnet-4-6 | 8,622 | 0 | 0 |
| claude-opus-4-7 | 4,468 | 0 | 0 |
| claude-haiku-4-5 | 1,038 | 0 | 0 |
| claude-opus-4-6 | 580 | 0 | 0 |

opus-4-7 (4.5k turns) and sonnet-4-6 (8.6k turns) had zero failures over large samples. This is not a usage-volume artifact — it is specific to opus-4-8.

Timeline: Days dominated by opus-4-7 (through ~May 28) had 0 failures. Failures begin when opus-4-8 became primary model (~May 30) and persist across CLI versions 2.1.156 and 2.1.160.

Characteristics:

• ~76% occur when generating a new tool call after a text segment (not mid-output)
• Preceding tool calls are small (Bash/Read/Edit, all <2KB input) — not caused by large outputs
• "(retry also failed)" is always present — the model re-breaks on retry
• Seen across multiple different projects/workdirs

What Should Happen?

claude-opus-4-8 should generate valid, parseable tool call JSON consistently — as opus-4-7 and sonnet-4-6 do (0 failures across thousands of turns). Turns should not abort with a parse error.

Error Messages/Logs

The model's tool call could not be parsed (retry also failed).

Steps to Reproduce

1. Use claude-opus-4-8 (select via /model)
2. Run any multi-step task involving tool calls (e.g. read a file then edit it)
3. Observe intermittently: turn aborts with "The model's tool call could not be parsed (retry also failed)"
4. Same task with claude-opus-4-7 or claude-sonnet-4-6 succeeds consistently

Note: No minimal repro — occurs stochastically at ~1.5% of turns. Statistical evidence from ~10k turns across multiple sessions.

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

N/A (CLI version not relevant — issue is model-specific: claude-opus-4-7 works, claude-opus-4-8 fails)

Claude Code Version

2.1.160

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

_No response_