Did Anthropic Secretly Nerf Claude Code? The Real Timeline
Anthropic quietly lowered Claude Code's reasoning effort and hit cache bugs that degraded output for weeks. Here's what actually happened.
Did Anthropic secretly nerf Claude Code?
Not exactly, but the effect on users was close enough that it barely matters. Anthropic quietly dropped Claude Code’s default reasoning effort from high to medium in March, hit a caching bug that made long sessions forgetful, and shipped a system instruction meant to shorten responses that measurably hurt coding performance. None of these were framed as downgrades when they shipped. All three were eventually fixed, but only after developers spent weeks complaining that Claude had gotten noticeably worse.
TL;DR
- Anthropic changed Claude Code’s default reasoning effort from high to medium in March, and later admitted this traded a small intelligence loss for lower latency and fewer usage-limit hits.
- A caching bug repeatedly stripped older reasoning out of long sessions, making Claude behave as if it had forgotten earlier context.
- A system instruction meant to shorten responses caused a measured 3% drop in coding performance before it was walked back.
- Anthropic says the API was never affected and denies deliberately weakening its models, so this is a product-layer story, not proof of a secret model swap.
- Claude Opus 5 launched with strong benchmark numbers but drew complaints from developers, including Theo Browne and Code Rabbit’s review testing, for being verbose and over-engineered on everyday tasks.
- Anthropic’s own documentation shows certain requests can be silently routed between Opus 5, Opus 4.8, and other models depending on internal safety classifiers, meaning two users on the “same” model can get different capability.
- None of this threatens Anthropic’s revenue in the short term, reportedly north of a $65 billion annualized run rate, but it does erode the trust that let Claude command premium pricing.
Other agents start typing. Remy starts asking.
Scoping, trade-offs, edge cases — the real work. Before a line of code.
What actually changed inside Claude Code?
Three separate issues stacked on top of each other over the course of several months, and each one independently made Claude Code feel less capable without any headline announcement.
The first was a change to reasoning effort. In March, Anthropic switched the default setting from high to medium. In practice, this meant Claude spent less computational effort “thinking” through a problem unless a user manually switched the setting back. Anthropic later confirmed this tradeoff explicitly: slightly lower intelligence in exchange for faster responses and fewer people hitting usage caps. That’s a legitimate product decision, but it was not surfaced to users as a capability change, so people noticed their tool getting worse before they understood why.
The second was a cache bug. Long coding sessions rely on the model retaining earlier reasoning and context. The bug repeatedly dropped older reasoning from those sessions, which made Claude appear forgetful and prone to repeating itself or re-solving problems it had already handled.
The third was a system instruction added to shorten Claude’s responses. Shorter answers sound like a usability win, but in practice it came with a measured 3% drop in coding performance, a concrete, quantified regression rather than a vague vibe shift.
All three were fixed. But the pattern, three separate quality regressions inside a matter of months, is what gave the “secret nerf” narrative its staying power, even though Anthropic pushed back hard on the framing.
Did Anthropic admit to any of this?
Partially. Anthropic acknowledged the reasoning-effort change and its tradeoffs directly. It also fixed the cache bug and reverted the response-shortening instruction after complaints. What the company disputes is the word “secret” in the sense of a deliberate, hidden downgrade of model intelligence to save on compute costs. Anthropic maintains the API was unaffected throughout, meaning developers building directly against Claude’s models rather than through Claude Code should not have seen the same degradation.
That distinction matters legally and technically. It does not fully matter to a user experience, though. If your daily driver is Claude Code and its default behavior gets quietly downgraded for weeks before anyone at the company explains why, the fact that it happened at the “product” layer instead of the “model” layer is a technicality that doesn’t restore lost work or lost trust. AMD AI director Stella Lorenzo was among the more prominent voices arguing exactly this, saying Claude could no longer be trusted for complex engineering work during the period these issues were live.
Is Opus 5 actually worse than it looks on benchmarks?
This is where the story gets more nuanced than a simple “nerf.” Opus 5 posted strong numbers on Anthropic’s own benchmarks, reportedly more than doubling Opus 4.8’s score on Anthropic’s frontier coding benchmark and tripling the next-best model on a separate reasoning benchmark, all while being cheaper than the prior flagship.
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
Real-world reaction diverged sharply from those charts. Theo Browne, developer and CEO of T3 Chat, described Opus 5 treating minor code comments like critical issues requiring massive rewrites. Code Rabbit’s controlled review testing found Opus 5 caught fewer known bugs than its production baseline while generating roughly four times as many low-value nitpicks.
The likely explanation isn’t fabricated benchmarks, it’s a calibration mismatch. Benchmarks reward a model for solving hard, well-defined tasks under ideal conditions. Everyday coding work is messier: incomplete instructions, interruptions, follow-up questions, and a preference for minimal, targeted fixes over sweeping rewrites. A model can genuinely improve at long autonomous builds and research tasks, which Opus 5 does appear to do well, while becoming more exhausting and less predictable in ordinary back-and-forth sessions. Code Rabbit’s own findings support this split, rating Opus 5 more useful as a builder than as a reviewer.
Does the same model name always mean the same experience?
Not necessarily, and this is arguably the most underappreciated part of the story. Anthropic’s own launch documentation states that certain cybersecurity-related requests inside Claude, Claude Code, or Claude Co-work can be silently routed away from Opus 5 down to Opus 4.8 by default. Other categories, including some safety and biology-related requests, get routed among Opus, Fable, and older models depending on an internal classifier.
Anthropic disclosed this system, so it isn’t secret in the strict sense, and the stated purpose is to keep harmless requests flowing instead of blocking them outright. But it does mean two customers typing the same prompt against the same selected model can receive different underlying capability without any visible signal that a substitution happened. Combined with Anthropic’s separate rollout of text watermarking that subtly influences word choice in outputs (which drew criticism from writers like John Gruber, though security researchers argue the quality impact is minimal), the overall picture is a product that can be reshaped underneath users in ways that are technically disclosed but practically invisible in the moment.
Is Claude Code still worth the premium price?
The pricing math has gotten harder to justify on brand loyalty alone. Opus 5 runs $5 per million input tokens and $25 per million output tokens, with the more restricted Fable 5 priced higher at $10 in and $50 out. Competing options have undercut that significantly: OpenAI’s promotional GPT 5.6 pricing sits at $2 in and $10 out, xAI’s Grok 4.6 at $2 in and $6 out, and Google’s Gemini 3.7 Flash at 75 cents in and $3.75 out.
Raw token pricing doesn’t capture the whole picture, since models differ in how much text they generate to solve the same problem. Code Rabbit’s testing found Opus 5 consumed roughly 50% more input tokens and 65% more output tokens than GPT 5.6 on equivalent review work, meaning the effective cost gap is wider than the sticker prices suggest. Anthropic has repositioned Sonnet 5 more competitively at $2 in and $10 out, and Opus can still be economical when it solves hard problems in fewer attempts. But the era where developers paid Claude’s premium without comparison-shopping appears to be ending. The Financial Times has reported that Fable 5 captured only around 11% of tracked US customer spending after launch, with cheaper, adequate alternatives limiting adoption even as Opus 5 performed better commercially.
Frequently Asked Questions
One coffee. One working app.
You bring the idea. Remy manages the project.
What is the Claude Code “reasoning effort” setting?
It controls how much computational effort Claude spends deliberating before answering inside Claude Code. Anthropic changed the default from high to medium in March, trading some intelligence for faster responses and fewer usage-limit hits, and later confirmed the tradeoff after user complaints.
Did Anthropic admit to intentionally degrading Claude’s models?
No. Anthropic acknowledged the reasoning-effort default change, the cache bug, and the response-shortening instruction, and fixed all three, but it denies deliberately weakening the underlying models and says its API was not affected by these issues.
Why did developers say Opus 5 felt worse despite strong benchmarks?
Benchmarks test defined tasks under ideal conditions, while everyday coding involves incomplete instructions and a preference for small, targeted fixes. Developers and testing firms like Code Rabbit found Opus 5 over-engineered simple problems and generated excessive nitpicks in code review, even though it performed well on long, autonomous, benchmark-style tasks.
Can Claude silently switch which model answers a request?
Yes, in disclosed circumstances. Anthropic’s documentation states certain cybersecurity, safety, or biology-related requests can be automatically routed between Opus 5, Opus 4.8, Fable, and other models based on internal classifiers, meaning the same prompt to the same named model can receive different underlying capability.
Is Claude Code still cheaper or more expensive than competitors?
Opus 5 and Fable 5 are priced higher per token than comparable offerings from OpenAI, xAI, and Google, and testing has shown Opus 5 can consume significantly more tokens to complete equivalent work, widening the real-world cost gap beyond list prices.