Claude Sonnet 5 – benchmark results

Claude Sonnet 5 arrives with a flashy resume: a top-tier intelligence score, support for text and images, and an enormous memory window big enough to swallow what feels like a small library. On paper, it looks like one of the smartest models around. In the comments, though, the vibe is less "wow" and more "wait, what am I even looking at?"

The loudest reaction was pure distrust. One commenter joked that so much of the benchmark page was missing or contradictory that it felt like the model itself had generated the report and then made half of it up. That line basically became the mood of the thread. People weren’t just debating whether Claude Sonnet 5 is good; they were side-eyeing the scoreboard itself.

Then came the cost-and-effort outrage. Multiple users complained that the model seems to burn through way more words, and therefore more usage, than rivals. In plain English: critics think it talks a lot, charges a lot of attention for it, and still runs into frustrating limits. One person flatly called it "yet another mediocre model," while another said they were tired of Anthropic’s caps and token-hungry behavior.

The spiciest drama was about real-world usability versus shiny test results. A commenter argued that if requests get blocked by provider rules for mysterious reasons, those should count as total failures in benchmarks, because that’s the actual user experience. Translation: the community isn’t just asking "Is it smart?" They’re asking "Can I actually use this thing without it eating my budget and refusing to work?"