Is Claude nerfed?

Deliberately? No evidence. Accidentally? Twice that we know of, for weeks at a time, with public postmortems afterwards. Anthropic’s position, stated a year apart in two separate incident reports, is that it never intentionally degrades its models. Its record is that when Claude did get worse, the cause was a bug, the detection took weeks, and the write-up was unusually candid.

So the honest answer to “is Claude nerfed?” is “not on purpose, but check the date.” Below is the record, then the checks.

August–September 2025: three infrastructure bugs

On September 17, 2025, Anthropic published a postmortem of three overlapping infrastructure bugs that intermittently degraded responses between August and early September:

• Aug 5 – Sep 18: a routing bug sent some Sonnet 4 requests to servers configured for the 1M-token context window. It started at 0.8% of requests and, after a load-balancing change on August 29, peaked at 16% of Sonnet 4 requests in the worst hour on August 31. Roughly 30% of Claude Code users had at least one message misrouted, and routing was “sticky”, so unlucky users stayed unlucky.
• Aug 25 – Sep 2: a misconfiguration on TPU servers occasionally assigned high probability to tokens that should almost never appear, producing Thai or Chinese characters in English answers and obvious syntax errors in code.
• Aug 25 – Sep 12: a change to token selection triggered a latent XLA:TPU compiler bug, confirmed for Haiku 3.5 and possibly affecting a subset of Sonnet 4 and Opus 3 requests.

The status page did not show an outage for any of this. The reports were, in Anthropic’s words, “difficult to distinguish from normal variation in user feedback” until late August.

March–April 2026: three product-layer changes in Claude Code

On April 23, 2026, Anthropic published an update on Claude Code quality reports tracing weeks of complaints to three changes in the product, not the model. The API was not affected.

• Default effort dropped from high to medium to fix very long latencies. Users read it as Claude getting dumber. Reverted April 7 after feedback; all users now default to xhigh on Opus 4.7 and high on other models.
• A caching “optimization” shipped March 26 with a bug: instead of clearing old reasoning once after an idle session, it cleared it on every turn for the rest of the session. Claude kept working with no memory of why it had done what it did, which surfaced as forgetfulness, repetition and odd tool choices. It also caused cache misses, which is why usage limits seemed to drain faster. Fixed April 10 (v2.1.101).
• A system-prompt line capping response length, added while tuning for the more verbose Opus 4.7, had an outsized effect on intelligence.

All three were resolved by April 20 (v2.1.116) and usage limits were reset for all subscribers on April 23. Press coverage at the time noted that Anthropic had initially implied little was wrong before the full write-up appeared.

The same claim, twice, in writing, a year apart:

“We never reduce model quality due to demand, time of day, or server load. The problems our users reported were due to infrastructure bugs alone.”Anthropic Engineering, September 17, 2025

“We take reports about degradation very seriously. We never intentionally degrade our models.”Anthropic Engineering, April 23, 2026

Take it for what it is: a public, on-the-record commitment that would be expensive to be caught breaking, made by a company that has twice published detailed evidence against its own competence rather than against its own honesty. It is not proof. It is a lot more than most “nerfed” threads have.

The thing Anthropic openly does adjust is how much you can use, and users routinely mistake that for how good it is. The clearest recent example: a 50% boost to Claude Code weekly limits ran from May 13 to September 13, 2026. When it ended, limits settled 25% above the original level, which is still a sixth less than the promotion. Anyone who noticed their week ending early on September 15 and concluded Claude got “nerfed” was wrong about the mechanism and right about the feeling.

Rate limits on the API work the same way: they cap requests and tokens per period, and the model behind them does not change. If your session is slower, shorter or interrupted, look at the meter before the model.

Five checks, in order, before you use the word:

1. Status page. status.claude.com shows incidents per product. Green does not mean fine (see above), but red means you can stop investigating.
2. The crowd. If it is an incident, other people are having it too. That is what the card below and the comments are for.
3. A control prompt. Keep one task you know Claude handled well last month. Run it fresh, same model, same settings. If it passes, your current conversation is the problem, not the model.
4. Settings and versions. Effort, thinking, model tier, app version. Write them down when things work so you can diff them when they do not.
5. The slope. One bad day is noise. Ten is a trend, and trends are what got both postmortems written.

“Claude” is a family, and half of all nerf arguments are two people comparing different members of it. As of this page’s review date, Anthropic’s model documentation lists Opus 5.5 as the recommended default for most workloads, Fable 5.1 for demanding reasoning and long-horizon agentic work, and Sonnet 5 as the current Sonnet, with older tiers still available. Apps choose among these on your behalf and change their choice without much ceremony.

When you vote, or post, or file a bug, name the model. “Claude was dumb today” is a feeling; “Sonnet 5 at medium effort in Claude Code v2.1.x was dumb today” is a report.