claude-code - 💡(How to fix) Fix [BUG] Opus 4.7 (1M context) severe latency regression starting evening of Apr 24, 2026 [4 comments, 2 participants]

Preflight Checklist

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

What's Wrong?

Claude Code with Opus 4.7 (1M context, xhigh effort) has become 10–15x slower than baseline starting the evening of April 24, 2026 (~19:00–20:00 UK time / ~18:00–19:00 UTC).

A simple /prime command on my project that has consistently taken ~30 seconds for the past 2+ weeks now takes 5–7+ minutes for the exact same operation, on the same project, with no changes on my side (same version, same config, same files, same model).

The output is correct and well-formed once it eventually completes — this is a pure latency issue, not a quality regression. Most of the wall-clock time is spent in the "Baking…" / "Clauding…" / "Crystallizing…" phase before any output appears.

Note: I'm on Claude Code v2.1.119, which is past the v2.1.116 fixes referenced in Anthropic's April 23 postmortem (https://www.anthropic.com/engineering/april-23-postmortem). This regression appeared after those fixes shipped, so this is likely a new/distinct issue.

What Should Happen?

/prime and similar lightweight commands (single small file read, ~150–200 input tokens) should complete in roughly 30 seconds, as they consistently did for the 2+ weeks prior to Apr 24, 2026.

Error Messages/Logs

No error messages — the session does not error or time out, it simply takes 5–7+ minutes to return a response. UI shows extended "Baking…" / "Clauding…" / "Crystallizing…" states with elapsed counters in the multiple-minutes range.

Steps to Reproduce

1. Open Claude Code in a project on macOS (Mac mini).
2. Confirm the model header reads: "Opus 4.7 (1M context) with xhigh effort · Claude Max".
3. Run any lightweight slash command that reads a small file. In my case: /prime, which reads a single ROADMAP.md (~161 input tokens) and produces a short status summary.
4. Observe elapsed time before completion.

Expected: ~30s. Observed (consecutive runs on Apr 25, 2026): 5m 13s, then 6m 58s.

Reproducible across:

• Fresh sessions (not a stale-session issue)
• Different terminals (VS Code integrated terminal and Apple Terminal)
• Different commands (not specific to /prime — also seen with general queries)

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

2.1.119 — same version. The version did not change. Was working at ~30s latency on this exact version up to the evening of Apr 24, 2026, then regressed without any local change. This is a backend regression rather than a CLI version regression.

Claude Code Version

2.1.119 (Claude Code)

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

VS Code integrated terminal

Additional Information

Suspected cause: Given the timing (post the v2.1.116 fixes from the Apr 23 postmortem) and that the symptom is specific to the Opus 4.7 1M context path, candidate causes include:

1. A backend serving issue specific to the 1M context variant — similar in shape to the 2025 misrouting incident where short-context requests were routed to 1M context servers (https://www.anthropic.com/engineering/a-postmortem-of-three-recent-issues).
2. A new regression introduced after v2.1.116.
3. Capacity / queueing on the Opus 4.7 serving path.

Related open issues (similar symptoms, not yet resolved):

• #50623 — Opus 4.7 performance degradation and excessive token consumption (Apr 19, still open)
• #49244 — Opus model quality regression starting ~Apr 15
• #37422 — Opus extremely slow inference (closed as dup, identical symptom shape)

Plan: Claude Max Project: small-to-medium codebase, well under 200k context — there is no legitimate reason for the 1M variant to be needed for these requests.

Happy to provide session feedback IDs, additional traces, or run diagnostic commands if useful.