Translated from the original Japanese article
Tech 7 min read

Claude Code's Quality and Quota Crumble on the Same Terrain

Claude Code Anthropic AI Agents Quality Usage Limits Prompt Caching
IkesanContents

Anthropic’s April postmortem broke down the Claude Code quality problem into three discrete bugs, while a DEV article attacked Max weekly quota exhaustion from a pure operational angle. Both landed in the same week but were read as separate stories. Split the same phenomenon into “quality” and “cost,” and the same infrastructure-side pressures show up in both. The HN thread (item?id=47878905) has the same two topics re-igniting in parallel, and it’s easier to follow them as one thread than two.

Sources for context:

Reframing the “same story” on two axes yields the following map. Quality degradation is orthogonal to changes that unintentionally strip internal state. Cost explosion, on the other hand, is observability blindness and setting-optimization failure happening simultaneously. What they share is that both assumed a design that doesn’t account for long-running-session behavior, and user experience broke along the way.

Overlaying the Timeline

The April 7–20 Anthropic fix sequence is, as information disclosure, high-quality work.

  • March 4: Default effort changed from high to medium (complaints surfaced during March)
  • March 26: An improvement meant to clear thinking after 1 hour idle was introduced with a bug that applied the clear every turn
  • April 7: All users reverted to xhigh/high
  • April 10: The 3/26 bug fixed
  • April 16: A verbosity-restriction system prompt added for Opus 4.7
  • April 20: Reproduced that this actually degraded coding performance; rolled back

Different origins, but the result was a single amplification on “compression, inference allocation, and missing observability” across the same session history. Proving Claude Code’s Quality Regression with 17,871 Thinking Blocks and 234,760 Tool Calls had already quantified the degradation as “a gradual decline appearing across February and March.” The current postmortem lines up neatly with that backdrop. The field observations in that article—edit accuracy degradation, repetition, a bias toward simplest—align with the 3/26 bug and the medium + high-thinking configuration.

Two Origins, One Shape of Damage

First, split them. The first axis is quality degradation: a user experience of “it got dumber,” driven by effort settings, thinking handling, and system prompt context. The second is consumption: cache, context, and long-running sessions producing cache miss chains that amplify observability opacity around actual usage.

Described separately it’s clean. In practice the two axes cross. The longer you run a session, the more thinking history and context you accumulate trying to preserve quality, and that bloat raises both recomputation cost and retry count. “Actions taken to recover quality” end up increasing cost, while “optimizations taken to reduce cost” force retries and simplifications that drop quality—the loop feeds itself.

On HN, Anthropic also explained that resuming an idle session triggers cache processing across all N messages, so with a ~900k history the 5-hour window can spike. That matches the user-facing experience: “I stepped away and came back, and cost became inexplicably high.” The explanation fills a CLI transparency gap. Before arguing “is there a problem or not,” the ordering of visibility needs a redesign, or the same pattern will recur.

Killing the Cost Boundary Hides the Quality Incident

As context grows in a long-running session, baseline per-tool-call cost doesn’t drop—it grows. The DEV article’s “2× billing feel at 100k context” is dangerous on quality grounds too. Changes that shallow thinking (effort reductions, thinking-history clears) and retry inflation both converge on the same “retry loop.” As Anthropic’s Unannounced Cache TTL Change and Claude Code Quota Exhaustion laid out, cache-TTL invisibility had already shown that a single 5-minute-window spec change could collapse users’ billing intuition overnight. Here again, settings changed while observation points lagged—same root.

Two outcomes:

  • A task that used to take 1 pass takes 4–6
  • Each recovery triggers the next quality regression

So the DEV recommendations (CLAUDE_CODE_AUTO_COMPACT_WINDOW, CLAUDE_CODE_DISABLE_1M_CONTEXT) aren’t pure cost-savers—they stabilize quality. Conversely, tightening thinking caps for cost reasons alone tends to inflate retries and drop quality. The effort tradeoff feeds back on itself.

Operational Priority: Observability → Setting Boundaries → Change Tracking

First check: observability. Aiming for both “high quality” and “low cost” while the contents of the usage bucket remain opaque is like twisting both steering handles without any dashboard info.

Next: setting boundary control. The DEV summary found CLAUDE_CODE_AUTO_COMPACT_WINDOW dropped to 200k, CLAUDE_CODE_DISABLE_1M_CONTEXT=1, and alwaysThinkingEnabled=false effective. These aren’t universal, but fixing “how much context we accumulate” alone suppresses both token bloat and long-session degradation.

Third: change tracking as a discipline. effort and system prompts shouldn’t be determined only by the default-value convenience of the moment. They should be pinned per task type in a shape that’s easy to A/B. Rather than pinning xhigh across the board, a per-task-difficulty map pays off more.

Concrete Implementation

First, Restore effort

For quality-first triage, the fastest move is staying at high or above and retrying on that baseline. As the postmortem showed, the one-directional swing to medium wasn’t even a latency win. Anthropic also reverted defaults on 4/7.

Then, Break Long-Running Session Design

Long-lived sessions are development-efficient, but when cache and context visibility is weak, quality-cost both degrade simultaneously. A pipeline running 8 hours a day easily involves 50+ tool calls per session. When context and tool calls self-propagate, quality and quota both hit their limits first. This is exactly the space How Compresr’s Context Gateway solves context exhaustion for AI agents addresses with a proxy-layer approach: separating “context the agent sees” from “context you’re billed for” is one viable answer.

Last, Revisit “Auto-Optimization” Requirements

alwaysThinkingEnabled, auto-updates, and overly aggressive cache settings raise cost and quality uncertainty if left at defaults. The most effective change in the DEV practice was alwaysThinkingEnabled: false plus explicit CLAUDE_CODE_AUTO_COMPACT_WINDOW—both highly reproducible in the field.

The “They Timed This Against GPT-5.5” Theory

The 4/24 postmortem + usage reset landed the same day OpenAI was known to be pushing GPT-5.5-related news. A nontrivial number of people on HN and X are speculating that Anthropic timed the release deliberately. Whether that’s true is unknowable, but at this moment—when “quality degradation” and “quota exhaustion” have both accumulated user distrust—dropping a reset as raw volume is certainly strong PR posture. The postmortem content itself is serious, but if you factor in the date-selection judgment, readers should separate “technical event” from “PR event” to avoid conflating the two.

This topic rewards stacking articles.

Reading in this order gives you “why the quality incident happened” → “why cost suddenly appeared to spike” → “why the boundary values are hard to explain” as a continuous line.