Rationing: The Postmortem

For weeks, Claude Code users watched their usage limits evaporate. Five-hour quotas burned in 90 minutes. Sessions generated hundreds of thousands of tokens nobody asked for. Anthropic said the tool was working as intended. Today, they published a postmortem saying it wasn't.

On April 23, Anthropic released a detailed postmortem identifying three separate product-layer bugs that caused Claude Code to consume users' quotas far faster than expected.¹ The company reset usage limits for all subscribers to compensate.²

The problems started on March 4. They weren't fully fixed until April 22.

Three Bugs, Seven Weeks

Bug 1: The reasoning downgrade. On March 4, a configuration change dropped the default reasoning effort from high to medium for paying subscribers on Opus 4.6 and Sonnet 4.6.³ Users paying $20 or $200 per month got a quieter, less capable version of the model they were paying for. Nobody was told.

Bug 2: The thinking cache wipe. On March 26, an optimization intended to clear stale "thinking" history from idle sessions had a critical flaw. Instead of clearing thinking history once after an hour of inactivity, it cleared it on every subsequent turn for the rest of the session.⁴ Claude progressively lost its own reasoning, producing repetitive output and wrong tool calls. This was fixed on April 10 in v2.1.101.

Bug 3: The verbosity prompt. A system prompt change made Claude generate more verbose responses, burning through token quotas faster. The postmortem does not specify when this was introduced or fixed, but it was reverted as part of the April 22-23 resolution.⁵

Instead of clearing thinking history once, it cleared it on every turn for the rest of the session.

What the Community Found First

While Anthropic was still diagnosing the problem, users were filing their own forensics. GitHub issue #38029 documented a session that generated 652,069 output tokens without any corresponding user prompts during a session resume. Estimated cost: $342.74.⁶

Issue #42338 identified that using --continue or --resume flags invalidated the entire conversation cache, forcing a full reprocessing of all prior context.⁷ A one-byte difference in tool attachment positioning was enough to break the prefix cache.

Issue #46829 tracked a separate regression: prompt cache TTL was reduced from one hour to five minutes around March 7.⁸ Anthropic acknowledged the change to The Register but disputed that it increased costs.⁹

Community researchers using reverse-engineering tools also identified a potential billing sentinel bug in Claude Code's runtime. That claim remains unconfirmed by Anthropic and is not addressed in the postmortem.¹⁰

The Fix

The April 22-23 fix included reverting the reasoning effort default back to high, reverting the verbosity prompt, and releasing v2.1.117 with a batch of bug fixes.¹¹ Anthropic also reset usage limits for all subscribers.

More notable was the process admission. From the postmortem:

"Going forward, Anthropic will implement broader evaluations for system prompt changes, ensure staff use public builds, improve their Code Review tool, and add soak periods for intelligence-affecting changes."

Read that again. They're committing to ensure staff use public builds. Which means, until now, they weren't required to. The people building the tool were not required to use the same version their customers were paying for.

The Pattern

This is the sixth entry in our rationing series. We've covered the initial quota complaints, Claude Code's removal from the Pro plan, the cache confusion, and the pay-per-overage model. The pattern repeats: users report a problem, the company says it's working as intended, and weeks later the company admits it wasn't.

This time, at least, there's a postmortem. That's progress. The question is whether the operational commitments hold. Broader evaluations, staff on public builds, soak periods for intelligence-affecting changes. These are table-stakes practices at any infrastructure company. That Anthropic needs to announce them as new commitments tells you where things were.

Disclosure

This article was written using Claude, an Anthropic product. The bugs described in this article directly affected the tools used to produce this publication. We have no relationship with Anthropic beyond being paying customers of their products. We believe honest coverage of the tools we depend on is more valuable than pretending we're neutral observers.