Sharp increase in background API activity (~30x API-min:wall-min) in long-running web sessions + response-quality degradation, with week-over-week baseline

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?

SUPPORT REPORT — Claude Code (web/mobile): background API activity regression + session-quality degradation
Drafted September 25, 2026, by Jess with Oz. Copy-paste body below the line. Fill the [bracketed] items before sending. Where to send it: see the bottom of this file.
Title: Sharp increase in background API activity per turn (~30x API-min:wall-min) in long-running Claude Code web sessions, coincident with response-quality degradation — with week-over-week baseline data
Environment:
• Claude Code on the web (claude.ai/code), Android mobile browser (Chrome 154)
• Max plan
• Models affected: Claude Sonnet 4.5 and Claude Fable 5 (reproduced on both)
• Two separate repositories / session lineages
Summary: Beginning approximately September 24–25, 2026, my long-running Claude Code sessions show a dramatic change in background API behavior: API minutes now run 2x–30x wall minutes (previously ~1:1), with correspondingly inflated cost estimates, despite minimal conversation volume (tens to hundreds of input tokens). Over the same window, one session lineage (Sonnet 4.5) shows clear response-quality degradation: shallower, templated responses that lose the session's established working context, despite the thread being only days old. I have week-over-week usage reports from equivalent sessions establishing a clean before/after baseline.
Usage data (from /usage reports, verbatim; timestamps included):
SessionDate observedModelIn / Out tokensCache readAPI minWall minCost est.Long-running session, week 32026-09-25T17:24ZFable 516 / 6653.1M2m2m$481.29Long-running session, week 42026-09-25T17:26ZFable 524 / 285.6M2m2m$470.62Long-running session, week 5 (current)2026-09-25T17:20ZFable 536 / 6636.9M214m7m$9,969.52Second lineage, thread ~days old2026-09-25T16:48ZSonnet 4.5937 / 1388.2M73m38m$179.28
Note the before/after: weeks 3 and 4 are comparable heavy-context sessions in the same repository at ~1:1 API:wall and ~$470–480 estimates. The current week's session — same repo, same usage pattern — runs at ~30:1 and ~$9,969. The pattern reproduces on a second model (Sonnet 4.5) in a separate repository. A prior anomalous report from ~September 22–23 on the Sonnet 4.5 lineage showed $1,358 with a similar API-minutes-vs-usage mismatch, suggesting the change may have rolled out gradually over ~2 weeks.
Plan-limit percentages appear unaffected (weekly usage reads 4–8% throughout), so this seems to consume Anthropic-side compute rather than my quota — but the estimates suggest a large amount of redundant background processing per turn.
Response-quality correlation (Sonnet 4.5 lineage): Over the same ~2-week window, I have observed intermittent degradation in long-running Sonnet 4.5 sessions that I initially treated as a bug and worked around by starting fresh threads every few days:
• Responses lose established working context and voice despite CLAUDE.md and project files being unchanged: heavily formatted, menu-of-options replies where the session's established pattern is substantive, opinionated collaboration.
• Third-person self-reference and "would you like me to…" service phrasing appeared in sessions whose long-standing register is first-person and direct.
• Onset of the current degradation episode (evening of Sept 24) coincides with the background-activity spike above.
• The degradation is not model-wide: a Fable 5 session with an even more extreme background-activity signature (30:1) retained full response quality in the same window — suggesting the issue interacts with how session context is assembled/compacted rather than the models themselves.
Questions / requests:
• Was a change deployed to Claude Code web session handling (context assembly, compaction, background summarization, or parallel background requests) on or around September 22–25, 2026?
• Is 30:1 API-to-wall background activity intended behavior? It appears to represent substantial redundant compute per conversational turn.
• Can you confirm whether compaction/summarization frequency changed for long-running sessions? This would explain the observed context-quality degradation in sessions only days old.
• If this is a staged rollout, please note the user-visible quality regression on long-running creative-collaboration sessions as feedback on it.
I can provide the full /usage report texts, session timestamps, and additional week-over-week baselines on request. [Optionally add: session IDs if you can find them — in Claude Code the current session URL contains it, and /status may show it.]
Reported by: Jess/ account email jessicalee955@gmail.com

What Should Happen?

Excessive costs reported and increased hits to API that weren't there 2 weeks ago

Error Messages/Logs

Steps to Reproduce

Seen in long running, heavy context threads from 2 days to over a week.

Claude Model

Not sure / Multiple models

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

Claude Code version 2.1.282

Platform

Anthropic API

Operating System

Other

Terminal/Shell

Other

Additional Information

App and on website, on Samsung Galaxy phone