# Why Is Claude Opus 5 Getting Bad Reviews Despite Top Benchmarks? > Source: > Published: 2026-09-13 00:00:00+00:00 # Why Is Claude Opus 5 Getting Bad Reviews Despite Top Benchmarks? Opus 5 tops Anthropic's benchmark charts but developers call it verbose and over-engineered. Here's the gap between scores and real coding work. ## Why is Claude Opus 5 getting bad reviews despite topping benchmarks? Opus 5 scores extremely well on Anthropic’s own benchmarks, including more than doubling Opus 4.8 on the company’s frontier coding test and tripling the next best model on a separate agentic benchmark. But developers using it day to day report something close to the opposite experience: a model that turns small fixes into large rewrites, buries simple answers in unnecessary explanation, and flags trivial code comments as urgent problems. The gap comes down to what benchmarks measure versus what daily coding work actually demands. ## TL;DR - Opus 5 posted strong results on Anthropic’s coding and agentic benchmarks, but developers including Theo Browne of T3 Chat describe it as verbose and prone to turning minor issues into unnecessary rewrites. - An independent review test from Code Rabbit found Opus 5 caught fewer known bugs than a production baseline while generating roughly four times as many low-value nitpicks. - Benchmarks reward completing a well-defined task with a clear success condition, while real coding sessions involve incomplete instructions, interruptions, and follow-up questions that reward restraint over exhaustiveness. - Anthropic has a documented history of quietly changing Claude’s behavior, including a reasoning-effort default that was lowered from high to medium and a caching bug that made long sessions forgetful, both fixed only after user complaints. - Opus 5 launch documentation confirms that certain sensitive requests can silently fall back from Opus 5 to Opus 4.8 or route to different models depending on internal safety classifiers, meaning two users on the same plan may get different underlying capability. - Opus 5 is priced at $5 per million input tokens and $25 per million output tokens, notably higher than competing flagship models from OpenAI, xAI, and Google, and Code Rabbit found it consumed far more tokens than a comparable GPT model on the same review task. - Reporting cited in the source material suggests Anthropic’s advanced Fable 5 model captured a modest share of tracked US customer spending after launch, indicating price and adequate cheaper alternatives are already shaping developer behavior. ## Plans first. *Then code.* Remy writes the spec, manages the build, and ships the app. ## What do the benchmarks actually show? Anthropic’s published results for Opus 5 are genuinely strong in specific categories. The model reportedly more than doubled its predecessor’s score on Anthropic’s frontier coding benchmark and achieved roughly three times the score of the next best model on an agentic benchmark referred to as ARC-AGI-3 style testing in the source coverage. Anthropic also positioned it as reaching near-flagship performance at a lower price than its highest tier model. On paper, this reads like an unambiguous upgrade. The trouble is that these benchmarks are built around a defined task with a clear success condition and a carefully chosen setup for the model to work within. That’s a controlled environment. It rewards a model that can grind through a hard, well-specified problem autonomously. It does not measure whether the model knows when to stop, whether it can distinguish a critical bug from a stylistic preference, or whether it respects the scope of what a user actually asked for. ## Why do developers say Opus 5 feels worse to use? The complaints center on three behaviors: verbosity, over-engineering, and poor judgment about scope. Theo Browne, developer and CEO of T3 Chat, described Opus 5 as treating every code comment like a critical issue demanding thousands of lines of new code. That’s a direct account from someone using the model in production, not a benchmark score. Code Rabbit ran a controlled review test comparing Opus 5 against a production baseline and found it caught fewer of the known problems the test was designed to surface, while generating around four times as many low-value nitpicks. That’s a meaningful signal: a model that produces more output isn’t necessarily producing more useful output. In code review specifically, noise has a cost. Every irrelevant flag is something a human has to read, evaluate, and dismiss. None of this means Opus 5 is a bad model overall. The same reporting found it performs well on long autonomous builds, research tasks, computer use, and visually complex projects, and that Code Rabbit rated it more useful as a builder than as a reviewer. The issue isn’t raw capability. It’s calibration: a model tuned to solve harder, more open-ended problems can become exhausting and unpredictable in the ordinary, low-stakes sessions that make up most of a developer’s actual day. ## Has Anthropic changed Claude’s behavior without telling users? ## Other agents ship a demo. Remy ships an app. Real backend. Real database. Real auth. Real plumbing. Remy has it all. Yes, in ways that are now documented rather than speculative. In March, Anthropic changed Claude Code’s default reasoning effort from high to medium, meaning the product spent less time “thinking” unless a user manually adjusted the setting. Anthropic later acknowledged this traded a small amount of intelligence for lower latency and fewer usage limit hits. Separately, a caching bug repeatedly dropped older reasoning from long sessions, making Claude appear forgetful and repetitive. A system instruction meant to shorten responses caused a measured 3% drop in coding performance. All three issues were eventually fixed, but only after weeks of public complaints, including from AMD AI director Stella Lorenzo, who said Claude could no longer be trusted for complex engineering work. Anthropic denies intentionally weakening its models and says its API was unaffected by these changes. That’s a fair distinction: quietly adjusting a default setting in a product isn’t the same as secretly nerfing the underlying model. But for a paying Claude Code user, the practical effect was the same either way. The assistant got less capable, without a changelog entry explaining why. ## Does the “Opus 5” label always mean the same model? Not necessarily. Anthropic’s own launch material states that certain cybersecurity-related requests inside Claude, Claude Code, or Claude Cowork can automatically fall back from Opus 5 to Opus 4.8, and that other safety-sensitive requests can be routed between Fable, Opus, and older models depending on an internal classifier. Anthropic disclosed this system openly, so it isn’t hidden. The stated goal is to avoid blocking harmless requests entirely while still applying extra scrutiny to sensitive ones. But the practical consequence is that two customers typing the same prompt into the same product tier can receive different underlying models without any indication in the interface. Combined with Anthropic’s announced text watermarking, which subtly influences word choice and drew criticism from writers including John Gruber, the pattern is one where “Opus 5” describes a routing system as much as a fixed model. Security researchers have argued the watermarking’s effect on output quality should be negligible, but the larger point stands: customers increasingly can’t be sure what they’re actually getting for their money. ## Is Opus 5 worth the price compared to competitors? Opus 5 costs $5 per million input tokens and $25 per million output tokens. Anthropic’s more restricted Fable 5 model costs $10 in and $50 out. For comparison, the source material cites OpenAI’s promotional GPT 5.6 pricing 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. Token prices alone don’t capture task cost, since a smarter model that solves a problem in fewer attempts can still be cheaper overall. But Code Rabbit’s testing found Opus 5 used roughly 50% more input tokens and 65% more output tokens than a comparable GPT 5.6 setup on the same review workload. That means the total bill is inflated by both the per-token rate and the sheer volume of text the model generates. Anthropic has made its Sonnet 5 tier more price-competitive at $2 in and $10 out, which suggests the company is aware of the pressure. But with capable, cheaper alternatives now available from multiple labs, developers have far less reason to simply accept Opus 5’s premium pricing on faith. ## Is Anthropic actually at risk of losing developer trust? The business numbers don’t show distress. Anthropic’s annualized revenue run rate reportedly exceeded $65 billion by the end of July, up sharply from the prior year, and the company is reportedly preparing for a possible public offering. Claude Code remains deeply embedded in professional workflows, and Opus 5 has real strengths in autonomous, long-horizon tasks. - ✕a coding agent - ✕no-code - ✕vibe coding - ✕a faster Cursor The one that tells the coding agents what to build. The risk is that trust erodes faster than revenue reflects it. If developers conclude that a model name doesn’t guarantee a consistent experience, that safety routing can quietly swap the model underneath them, and that competitors offer adequate performance at a fraction of the cost, they can shift workloads gradually and quietly, long before it shows up in a quarterly report. Reporting cited in the source material notes that Anthropic’s Fable 5 model captured only a modest share of tracked US customer spending after launch, which suggests this shift may already be underway in pockets of the market. ## Frequently Asked Questions ### Why does Opus 5 score well on benchmarks but perform poorly in daily use? Benchmarks test a model’s ability to complete a well-defined task with a clear success condition, often in a carefully constructed agent setup. Daily coding work involves ambiguous instructions, interruptions, and the need to judge when a small fix should stay small. Opus 5 appears tuned for the former, which can make it verbose and over-engineered in the latter. ### Did Anthropic secretly make Claude worse at any point? Anthropic changed Claude Code’s default reasoning effort from high to medium, which it later admitted reduced intelligence slightly in exchange for speed. A separate caching bug also degraded long-session performance. Anthropic denies intentionally weakening its models and says its API was unaffected, but both issues affected paying users before being fixed. ### Can Opus 5 route to a different, weaker model without telling the user? Yes, in specific documented cases. Anthropic’s own launch material says certain cybersecurity and safety-sensitive requests can fall back from Opus 5 to Opus 4.8 or route to other models based on internal safety classifiers. This is disclosed publicly but not visible in the moment to the user making the request. ### Is Opus 5 more expensive than competing models? Yes. Opus 5 is priced at $5 per million input tokens and $25 per million output tokens, higher than comparable flagship pricing from OpenAI, xAI, and Google cited in industry coverage. Testing by Code Rabbit also found Opus 5 used significantly more tokens than a comparable GPT model on the same task, increasing real-world cost further. ### Is Claude Opus 5 a bad model overall? No. It shows genuine strength in long autonomous builds, research, computer use, and visually complex projects. The criticism is concentrated on everyday coding and code review tasks, where its tendency toward verbosity and unnecessary rewrites works against the kind of quick, targeted help developers expect.