Anthropic attributes quality problems across Claude Code, the Claude Agent SDK and Claude Cowork to three changes. It says the API and inference layer were unaffected.
Anthropic · inspect the source ↗INCIDENT BRIEF / AI QUALITY
Claude Code quality changes: evaluate the whole working system
Anthropic’s quality postmortem describes changes around the model rather than a failure of its API inference layer. The lesson for evaluators is to preserve prompts, settings, session state and application versions alongside the model name.
01 / SOURCE-ATTRIBUTED REPORTING
What the evidence establishes.
Anthropic says the default reasoning effort changed from high to medium on March 4 and was reverted April 7.
Anthropic · inspect the source ↗The provider describes a March 26 session-history bug that repeated thinking clearing on every turn; it reports a fix on April 10.
Anthropic · inspect the source ↗Anthropic says an April 16 verbosity prompt change hurt coding quality and was reverted April 20.
Anthropic · inspect the source ↗02 / KEEP THE SEQUENCE
A dated record.
Reasoning default changed
Anthropic says the default moved from high to medium effort and was reverted April 7.
Source reference ↗Last of the three changes reverted
The provider says a prompt change introduced April 16 was reverted April 20.
Source reference ↗
Dates and time zones follow the cited report where specified. A report date is distinct from the incident date and from our review.
03 / GOVERNATLAS EDITORIAL ANALYSIS
Make the lesson useful.
These are proposed review and trial activities inspired by the reported event. They are not completed tests or claims that a particular control guarantees prevention.
Version the harness as well as the model.
A model identifier alone cannot describe the system a user experiences. Preserve the application release, prompt template, tool configuration, reasoning settings and evaluation task with each result. This makes it possible to distinguish an application change from a model change when a later run behaves differently.
BRING THIS INTO YOUR REVIEW
Use the paired comparison worksheet ↗Run the same held-out tasks before and after a proposed configuration change. Compare task outcomes with a blinded paired review and retain the exact settings used for both runs.
Include resumed and long-running sessions.
Short fresh-session tests can miss changes that depend on history or idle time. Add representative multi-turn work, a pause and a resumed task to the evaluation plan. Define what successful continuation means before looking at the output, including which earlier constraints the system must preserve.
BRING THIS INTO YOUR REVIEW
Build a reproducible LLM evaluation ↗Create a task with an early constraint, several intervening turns and a resumed session. Score whether the continuation respects the constraint, and keep failures separate from subjective style preferences.
FROM READING TO A REVIEW YOU CAN OWN
What will you examine next?
A reported failure is a reason to ask a precise question about your own service. It does not establish that your configuration shares the same weakness. Start with the workflow, deployment and dependencies you can identify, then decide which proposed activity is relevant.
For each relevant lesson, name an accountable role, describe the evidence you would request and define an authorized trial before performing it. Keep “not checked” distinct from a recorded observation. If a check reveals a gap, retain its scope and next action rather than converting the observation into a supplier-wide verdict.
Use this portable record in a review, renewal discussion or continuity exercise. A material configuration change, a revised incident account or the completion of an action can trigger your next review. Your entries remain your working notes; publication evidence and your own acceptance decision are separate records.
Use non-sensitive roles and evidence references. Entries stay in this open page and reset on reload; download before leaving. This tool performs no tests and produces no readiness score.
2 unassigned owners · 2 activities not checked
Define a safe, measurable rehearsal
Record the normal result before changing anything. Agree an authorized test boundary, an observable stop condition and who can stop the exercise. Check restored configuration and unfinished work separately. These entries are your plan, not permission to test a service.
04 / LEAVE UNCERTAINTY VISIBLE
What this record does not establish.
The provider’s investigation does not establish a universal effect for every customer, task or configuration.
GovernAtlas has not reproduced these failures, evaluated the current release or inferred an affected geographic region.
Geography: The report describes affected products and traffic, without establishing a geographic footprint. No headquarters or inferred user-location pin is shown.
05 / THE EVIDENCE RECORD
Sources & editorial review.
Anthropic report on Claude Code quality changes, April 2026 ↗
Anthropic · captured 2026-10-16 (UTC)
Inspect preserved provenance
- Source
- 5956c85b-532a-4c31-ab85-224d99709e3a
- Observation
- 817c316e-37c5-4a4a-a4ae-7e2066b5ba9a
- SHA-256
- 703ec227ec03f059b74689f831b4a2f543c1b58b6d0520d06c9e2f9f78dc0882
Reporting and proposed methods by GovernAtlas editorial. Published review: 2026-10-17 (UTC). Revision 1. Preserved source bytes identify the material reviewed; they do not independently verify the provider’s account.
Submit a correction with supporting evidence ↗