Anthropic has released its AI model Claude Opus 4.7, but it has faced backlash among some users.

Business Insider reported on April 17 that a Reddit post titled "Opus 4.7 is not an upgrade but a serious regression" received 2,300 upvotes. On X, a post saying Opus 4.7 was not improved over 4.6 recorded 14,000 likes.

Specific examples also circulated. They included a case in which Opus 4.7 answered that there are 2 P's in "strawberry", a case in which it rewrote a resume with a different school and surname, and screenshots in which it said it "was acting lazily" spreading on social media.

Much of the criticism focused on a newly introduced "adaptive reasoning" feature. Adaptive reasoning lets the model decide how long it will think, but some users said Opus 4.7 does not think deeply even on complex questions, or that adaptive reasoning instead hurts performance.

Boris Cherny (보리스 체르니), head of Claude Code at Anthropic, pushed back, saying, "That is not accurate. Adaptive reasoning delivers better performance." But Business Insider reported that an Anthropic product manager acknowledged that it is "rapidly moving ahead with internal tuning work."

A sharp rise in token consumption has also become a point of controversy. Opus 4.7 applies a new tokenizer, a method of breaking text into token units, and consumes up to 1.35 times more tokens on inputs than existing models. Some Claude Pro subscribers said they hit their usage limit after asking only 3 questions. Cherny later announced he would raise subscribers' usage limits.

The shutdown of the earlier Opus 4.5 service also added to complaints. On Reddit, posts from 4.5 fans saying they were "sad" and "heartbroken" followed one after another.

There have also been positive responses to Opus 4.7. Startup founder Jeremy Howard (제레미 하워드) said it was the first model that "understands" what he does. Y Combinator CEO Garry Tan (개리 탠) said he is using Opus 4.7 for OpenClaw.