Anthropic's release of Claude Opus 4.7 has landed with a thud rather than a splash, as users across Reddit, Discord, and developer forums raise pointed questions about whether the company has quietly repackaged a nerfed version of its previous model at a steeper effective price.
The anger is loud and, by most measures, warranted. Within days of the April 2026 rollout, the AI community coalesced around a damning narrative: Opus 4.7 is not a meaningful upgrade over Opus 4.6. It is, many believe, the pre-nerf build of 4.6 dressed in a higher model number, slipped out the door with a tokenizer change that functions as a stealth price hike. That tokenizer now consumes up to 35% more tokens for identical inputs, which means users are paying significantly more per conversation without any corresponding gain in capability. Anthropic has not addressed this directly, and the silence is doing nothing to calm the crowd.
Usage limits have become a flashpoint of their own. Subscribers across every tier are reporting that they're hitting their 5-hour and weekly caps after only a handful of prompts, a problem that didn't manifest with the same severity under 4.6. For professional users who depend on sustained, high-volume workflows, this isn't an inconvenience. It's a productivity collapse. The pattern is consistent enough across reports that it's difficult to dismiss as anecdotal.
The most technically damaging accusation involves long-context retrieval. On the MRCR benchmark, a standard measure of how well a model can locate and reason over information buried deep in a long document, Opus 4.7 scores 32.2%. Opus 4.6 scored 78.3%. That is not a minor regression. That is a collapse. For anyone using Claude in legal document review, financial analysis, or research synthesis , tasks where long-context fidelity is the entire point , this number is alarming. An Anthropic developer responded to the criticism by explaining the company is phasing out the MRCR benchmark, a response that struck most observers as a non-answer rather than a reassurance. You don't retire a benchmark because your model suddenly performs well on it.
Claude Code, Anthropic's flagship developer product, has its own crisis. Multiple engineers report that 4.7 is flagging routine, benign code as malware and refusing to complete basic edits. A model that won't write code is not a coding assistant. The false positive rate appears high enough that some developers have already reverted to earlier model versions or switched tooling entirely. This is a particularly bad look given how aggressively Anthropic has marketed Claude Code as an enterprise-ready product.
Adaptive Thinking and the disappearing toggle
The removal of the Extended Thinking toggle from the web interface has landed poorly with power users who relied on it to manage cost and latency. Its replacement, Adaptive Thinking, makes the reasoning depth decision for you. Some users appreciate the simplicity. Most who cared about the original toggle did not want simplicity , they wanted control. Framing an autonomy reduction as a feature upgrade is a difficult sell, and Anthropic is finding that out in real time.
Then there is Mythos. Anthropic's unreleased flagship model has appeared in benchmark comparison charts alongside 4.7, and the community reaction has been nearly uniform: eye-rolls. In areas where Opus 4.7 trails competitors, including GPT-5.4 in several reasoning and instruction-following benchmarks, Mythos floats nearby in the chart looking capable and impressive. The effect, whether intentional or not, is to redirect attention from 4.7's actual performance toward a model nobody can use yet. Doing this once might be forgivable. Making it a pattern turns goodwill into skepticism fast.
What makes this moment particularly consequential is that Anthropic built its reputation on a specific promise: that it was the careful, trustworthy alternative to the move-fast operators in this space. That positioning has real commercial value, especially with enterprise clients making multiyear infrastructure commitments. A stealth tokenizer change, a benchmark abandonment, a malfunctioning code product, and a governance-of-thinking-depth controversy all arriving simultaneously erodes exactly the kind of trust that brand was built on.
The AI model market in 2026 is genuinely competitive in a way it wasn't even twelve months ago. OpenAI, Google DeepMind, and several well-funded challengers are all shipping at pace. Users annoyed enough to leave now have real places to go. Anthropic should watch its churn numbers very closely over the next few weeks, because community sentiment this negative has a way of becoming enterprise procurement sentiment before long.
Also read: Anthropic's Opus 4.7 is drawing rare unanimous criticism from power users who say the model has lost its spark • Google's Gemini Pro couldn't draw a map of Europe and the internet has opinions about why • Anthropic's Claude Opus 4.7 posts a jarring benchmark regression that has enterprise AI teams asking uncomfortable questions
