Opus 4.7 has been released and tested to be indeed strong. But I found that when it was released 4.6 two months ago, it was listed in the official benchmark table as beating GPT-5.4, "currently one of the strongest". In today's 4.7 comparison table, the same 4.6 suddenly falls behind GPT-5.4. The model has not changed, the evaluation set has not changed, only the comparative perspective has changed. To put it simply, every time a new model is released, the previous generation is quietly "recalibrated" in order to highlight the improvements of the new version. Not only Anthropic, OpenAI and Google play like this every time they have a press conference. It's better to look at third-party long-term reviews and real user feedback than to look at model releases and benchmarks. The official table always shows that the new model wins