Opus 4.8: Frequent silent empty turns / 'tool call was malformed' errors (workaround: switch to 4.7)

Summary

In Claude Code with Opus 4.8, we encountered frequent failed turns that surface as either of two modes:

• Silent failure: the "Cooked for Ns" indicator runs for ~15-30s, then the turn ends with no assistant text and no visible tool call delivered to the user.
• Visible error: a system message Your tool call was malformed and could not be parsed. Please retry. is shown. The automatic single retry does not converge.

Symptoms persisted across multiple sessions over two consecutive days (2026-06-13 → 2026-06-15) on the latest CLI build.

Environment

• Claude Code CLI: 2.1.177 (verified via claude update)
• Model: claude-opus-4-8 (1M context)
• Platform: macOS (Darwin 25.5.0)
• Shell: zsh

Observed Correlation (anecdotal)

In our sessions, the failure appeared to cluster around turns where the assistant wrote one or more sentences of preamble text and then attempted to invoke a tool such as WebFetch or Bash. Turns that began directly with a tool call were less likely to fail.

This is an anecdotal observation, not a controlled test. We did not isolate confounders such as language (the affected sessions were primarily in Chinese), tool type, or prompt length. Reporting it in case it correlates with something the team is already tracking.

Informal Reproduction

In Claude Code with --model claude-opus-4-8, the following prompt shapes reliably produced failed turns for us:

• "Go check status.claude.com and tell me if Opus is having issues right now."
• "Look up the latest update on the Fable 5 / Mythos 5 suspension."

Outcome over ~10 attempts: a mix of successful turns, silent empty turns, and the explicit tool call was malformed error. Not deterministic, but failure rate was high enough that 3-4 blank turns in a row occurred multiple times.

Workaround

• /model claude-opus-4-7 resolves the symptom. No failed turns of either mode observed on 4.7 across the same workload in the same session.
• On 4.8, instructing the model to "call the tool directly without writing any introduction first" appeared to reduce the failure rate but did not eliminate it.

Status Page Discrepancy

status.claude.com marked "Elevated errors on Claude Opus 4.8" as resolved at 2026-06-13 02:11 UTC. However, the symptoms described above continued to occur on the same CLI build through 2026-06-15, after that incident was marked resolved. We suspect the resolved incident addressed a different elevated-error wave rather than the cause behind these failed turns.

Impact

• Two days of degraded experience with the flagship model in a real workflow.
• The silent-failure mode is particularly disruptive: users cannot distinguish "model is thinking" from "turn produced nothing usable." Many users will assume their machine, network, or prompt is at fault.
• Forces fallback to Opus 4.7 for any tool-heavy workflow.

Asks

• Confirm whether the cause is known and tracked internally; if so, please consolidate. We do not know whether the silent empty-turn mode and the explicit tool call was malformed error share a root cause.
• Surface tool-call failures to the user immediately rather than allowing a silent empty turn — currently the UX makes the model look hung when it is actually done.
• If the underlying issue remains open, consider reflecting that on status.claude.com as distinct from the resolved elevated-error incidents.