Summary (TL;DR)
Opus 4.7 uses a new tokenizer that boosts input tokens by up to 35% and defaults to xhigh effort, increasing total token spend. Adaptive thinking is now the only thinking mode, and reasoning summaries are hidden unless you opt in. Many users report the model is dumber, burns through subscription limits faster, and refuses more tasks due to new cyber safeguards. Some find quality on par with 4.6 after disabling adaptive thinking and raising effort. The launch follows weeks of complaints about 4.6 being silently degraded. Anthropic says prompts need retuning for stricter instruction following. Competitor Codex is gaining users frustrated with Claude's limits and opacity.
Public access expired
Save this link to your readplace queue and read every link without expiration.
Save to My Queue