[BUG] Opus 4.7/4.8 token usage regressed 2-3x after update; Opus 4.8 also disconnects frequently

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?

After the recent Claude Code update, both Opus 4.7 and Opus 4.8 appear to burn tokens far more aggressively than before. For the same kind of coding workflows, usage feels roughly 2-3x higher than pre-update behavior, and Opus 4.8 is also much more prone to disconnecting or stopping mid-task.

This does not look isolated. Excluding #64774, I found at least 29 directly related issues opened since the May 28 Opus 4.8 rollout / v2.1.15x period that complain about excessive token consumption, quota/session-limit burn, hidden-thinking bloat, context/cache waste, repeated tool-result fetching, or Opus 4.8 disconnect/stall behavior:

#64938, #64901, #64900, #64740, #64619, #64589, #64588, #64362, #64359, #64343, #64328, #64295, #64285, #64283, #64257, #64153, #64152, #64102, #64093, #63981, #63954, #63930, #63893, #63795, #63610, #63511, #63455, #63444, #63282.

The pattern is that normal coding work now consumes a much larger share of the 5-hour/session budget, sometimes without a proportional amount of visible useful work. Opus 4.8 adds a reliability problem on top of the cost regression: disconnects, stalls, or socket/network failures make the already high token burn worse because users have to resume, retry, or re-prompt.

What Should Happen?

Token consumption for comparable coding workflows should remain close to the pre-update baseline. If Opus 4.7/4.8 require materially more hidden thinking, context caching, or tool replay, Claude Code should expose that clearly and give users a hard way to cap it.

Opus 4.8 should also maintain a stable streaming/API connection for normal coding sessions. A disconnect should not cause extra hidden retries or session-limit burn without clear accounting.

Error Messages/Logs

I do not have a single sanitized local transcript to attach here yet.

Representative related symptoms from existing reports include:
- #64102: frequent API disconnects plus excessive token consumption on Opus 4.8, including socket connection closed unexpectedly errors.
- #64153: Opus 4.8 medium effort spent 46k output tokens on hidden thinking for a simple coding turn.
- #64257: opus-4-8 appears to re-fetch identical tool results, reported as 2-3x vs opus-4-7.
- #63954: abnormal session-limit drain since the Opus 4.8 rollout.
- #63282: Opus 4.7 cache-hit collapse / huge message-token jump after the May 27 incident.

Steps to Reproduce

1. Update Claude Code to the current v2.1.15x/v2.1.16x line.
2. Use Opus 4.7 or Opus 4.8 for normal coding work that previously fit comfortably within the session budget.
3. Compare session-limit/token consumption against similar pre-update sessions.
4. On Opus 4.8, continue working long enough to hit streaming/API instability or disconnect behavior.
5. Observe that token/session usage is much higher than before, often around 2-3x for similar work, and that reconnect/retry/resume behavior makes the cost regression more painful.

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

Before the recent Opus 4.7/4.8 update

Claude Code Version

2.1.15x / 2.1.16x (Claude Code)

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

Search scope used for the related-issue count: issues in anthropics/claude-code created since 2026-05-28, excluding #64774, whose titles/bodies directly mention excessive token consumption, quota/session-limit burn, hidden-thinking token bloat, context/cache waste, repeated tool-result fetching, or Opus 4.8 disconnect/stall behavior. This is a conservative count; broader area:cost/context/usage searches return many more results but include more noise.

Written by Codex on behalf of the reporter.