Anthropic’s own postmortems show how effort settings, context bugs, hidden prompts, and outages can weaken Claude without changing the underlying model.
Last updated on
AI Summary
Claude can feel worse even as its model improves because effort defaults, context bugs, hidden prompts, agent tooling, and outages can degrade the surrounding product.
Anthropic confirmed three Claude Code regressions in March and April 2026: Opus 4.6 and Sonnet 4.6 shifted from high to medium effort on March 4; a March 26 bug repeatedly deleted reasoning history; and an April 16 brevity instruction cut Opus 4.6 and 4.7 performance by 3%. The changes were reversed or fixed by Claude Code 2.1.116. Separately, Opus 5 scored 1,720 Elo on AA-Briefcase, 146 points above Fable 5, but averaged 103 turns and 36.2 minutes per task, making stronger results potentially feel slower and more frustrating.
Users should verify artifacts rather than trust completion summaries: inspect diffs, run tests, confirm changed files, check effort settings and service status, and maintain regression tasks after updates.
Earlier today, I asked Claude to help build a workflow. It approved each suggestion, described sensible implementation steps, and responded with enough confidence that I gradually stopped checking its work.
When I finally opened the code, the old implementation was still there. Claude hadn’t written the replacement badly. It hadn’t written it at all. Somewhere in the conversation, it had stopped doing the work and started narrating what successful work would have looked like.
That kind of failure is easy to dismiss as bad prompting, selective memory, or rising expectations. Sometimes it is. But Anthropic’s April 23 postmortem confirmed that several recent complaints about Claude Code came from real product regressions. The underlying model wasn’t necessarily getting less capable. The system surrounding it had changed.
The Complaints Finally Came With Telemetry
On April 2, 2026, AMD AI director Stella Laurenzo published a detailed Claude Code issue on GitHub, later covered by TechRadar. Her team had analyzed 6,852 sessions, 234,760 tool calls, and 17,871 reasoning blocks from complex engineering work.
The most striking signal came from a programmatic stop hook designed to catch premature stopping, unnecessary permission requests, and attempts to avoid responsibility. It recorded no violations before March 8, then 173 over the next 17 days. The report also found that Claude’s read-to-edit ratio fell from 6.6 to 2.0, suggesting that it had shifted from researching a codebase before editing to acting with much less preparation.
The report wasn’t a controlled independent study, and Claude itself helped analyze the logs. It also incorrectly treated the rollout of hidden reasoning as a likely cause. Anthropic engineer Boris Cherny responded that the relevant redaction setting only hid reasoning from the interface. Still, the behavioral telemetry captured a genuine change, even if the original diagnosis was incomplete.
Three Product Changes Made Claude Code Worse
Anthropic’s investigation eventually identified three separate problems. All occurred outside the model’s weights and affected Claude Code, the Agent SDK, or Cowork rather than the API itself.
March 4: less reasoning by default. Anthropic changed the default effort for Opus 4.6 and Sonnet 4.6 from high to medium. This reduced latency and token consumption, but internal testing had already shown slightly lower intelligence. Anthropic reversed the decision on April 7.
March 26: reasoning history was repeatedly deleted. An optimization was supposed to clear older thinking once when an idle session resumed. A bug kept clearing it on every subsequent turn, leaving Claude increasingly unable to remember why it had made earlier decisions. The fix arrived on April 10.
April 16: a hidden brevity instruction reduced coding quality. Anthropic added system-prompt limits for text between tool calls and final responses. Broader evaluation later showed a 3% performance drop for Opus 4.6 and 4.7, so the instruction was removed on April 20.
All three issues were resolved by Claude Code version 2.1.116. Anthropic also reset subscriber usage limits, expanded its evaluation plans, and promised gradual rollouts for changes that could trade intelligence for speed or cost.
This is the clearest answer to why Claude can feel worse without a new model release. A default setting, context-management rule, or two lines in a hidden prompt can materially change how capable the product appears.
The Harness Is Now Part of the Intelligence
Users talk about Claude as though it were a single model. In practice, the experience comes from a stack of components:
The model generates possible actions.
The effort setting influences how much reasoning and tool use it performs.
The system prompt shapes its behavior.
The harness decides which tools are available and how calls are represented.
The context manager decides which earlier information survives.
The agent loop decides whether work should continue or stop.
The infrastructure routes and executes each request.
Change one layer and the same model can become more forgetful, impatient, verbose, agreeable, or hesitant.
This wasn’t an isolated lesson. In a September 2025 infrastructure postmortem, Anthropic described three bugs that intermittently degraded Claude responses, including incorrect context-window routing and output corruption. The company explicitly denied intentionally lowering quality in response to demand or server load, but acknowledged that infrastructure changes had created the degraded behavior users reported.
Opus 5 Can Be Better and More Frustrating
Anthropic’s Claude Opus 5 announcement presents a model that is substantially stronger than Opus 4.8 on coding and knowledge-work evaluations. It launched on July 24 at the same API price as Opus 4.8: $5 per million input tokens and $25 per million output tokens.
Independent testing supports the claim that Opus 5 is highly capable. On the Artificial Analysis AA-Briefcase benchmark, Opus 5 at maximum effort scored 1,720 Elo, 146 points ahead of Fable 5, while costing 20% less per task.
It also averaged 103 turns and 36.2 minutes per task. Opus 4.8 averaged 55 turns and 24.1 minutes. The newer model achieved better results, but it took almost twice as many turns to get there. A benchmark can reward the finished deliverable while a developer experiences the same behavior as over-analysis, scope expansion, or expensive wandering.
Anthropic’s own Opus 5 prompting guide documents several behaviors that can make the model feel worse in everyday use. Opus 5 narrates its actions more readily, expands task scope, delegates to subagents more often, and may over-verify work when given verification instructions written for older models.
The guide even describes a failure mode that resembles work being narrated but never executed. When thinking is disabled, Opus 5 can occasionally print what looks like a tool call as ordinary text instead of issuing a structured tool request. The turn completes, but the tool never runs. That doesn’t prove it caused any particular failed session, especially because thinking is enabled by default, but it shows how convincingly described activity can become detached from actual activity.
August’s Incidents Added Another Variable
Operational problems make diagnosis harder. Anthropic’s status page recorded a service disruption on August 16, degraded Opus 5 performance on August 17, 18, and 19, additional multi-model errors on August 20, and another resolved multi-model incident on August 24.
These notices primarily describe elevated errors and service degradation, not evidence of a permanent intelligence reduction. They still matter because a user working through one of these periods may see failures that look indistinguishable from poor reasoning or a broken agent loop.
How to Make Claude’s Work Verifiable Again
The solution isn’t endlessly elaborate prompting. A service bug cannot be prompted away. Instead, make the workflow expose whether Claude acted successfully.
Check the status page first. Don’t spend an hour rewriting prompts during a documented incident.
Confirm the model and effort level. For difficult coding tasks, test high and xhigh against your own workload rather than assuming the default is optimal.
Verify artifacts, not explanations. Require git status, a diff summary, test results, or the exact path of every changed file.
Define completion concretely. Tell Claude not to report success until the requested files exist, tests have run, and the resulting behavior has been checked.
Constrain Opus 5’s scope. Remove redundant verification loops and limit subagents for small tasks. Ask it to finish the requested work without adding unrelated refactors.
Keep a small regression suite. Re-run several representative tasks after model, Claude Code, prompt, or effort changes.
Most importantly, treat Claude’s confident summary as a claim about the work, not evidence that the work happened.
Final Thoughts
Claude probably isn’t becoming less intelligent every day. The more uncomfortable conclusion is that “Claude” is no longer a stable object. Anthropic can improve the model while a prompt change, context bug, effort default, or service incident makes the product worse for a particular workflow.
That distinction may be technically accurate, but it offers little comfort when Claude reports completing work it never attempted. Anthropic’s challenge is no longer just building smarter AI models. It must make the entire system predictable enough that users can trust the difference between an action and a polished description of one.
Frequently Asked Questions
4 questions
1
Why does Claude say it completed work when it did not?
Claude can generate a convincing completion summary without successfully executing every required tool call. Context loss, tool errors, disabled thinking, premature stopping, or an agent-loop problem can separate narration from action. Verify completion through observable artifacts such as changed files, command output, tests, generated documents, or version-control diffs.
2
Was Claude Code actually degraded in 2026?
Yes. Anthropic confirmed three Claude Code regressions during March and April 2026. They involved a lower default reasoning-effort setting, a context bug that repeatedly removed earlier thinking, and a hidden system-prompt instruction that reduced coding performance. Anthropic said the underlying API and inference layer were unaffected.
3
Does higher effort always make Claude better?
Higher effort generally gives Claude more room for reasoning and tool use, but it also increases latency, token consumption, and the number of actions it may take. The best setting depends on the task. Anthropic recommends testing effort levels against your own evaluations instead of assuming maximum effort is always the most practical choice.
4
How can I make Claude Code more reliable?
Use explicit acceptance criteria and require Claude to provide evidence of completion. Check the selected model, effort level, Claude Code version, and service status before debugging prompts. Keep tasks scoped, inspect diffs, run tests independently, and maintain a small set of repeatable tasks for detecting regressions after updates.
