Opus 5.5 'nerf' reports: trackers say it's normal, check effort first
The effort default quietly fell to medium, so it reasons less and feels weaker — yet a drift benchmark finds no real decline. Plus Fable 5.2's record run.

Copy markdown
The reports — and the benchmark watching daily
Builders are flagging Opus 5.5 'thinking less' and slipping on hard tasks since early November. livenerf — a pre-registered benchmark that reruns a frozen 78-question panel every day on a Claude Max seat via headless Claude Code — so far reads within its normal range, and Anthropic hasn't confirmed any change.
Check your effort setting before blaming the model
Opus 5.5 shipped with its effort default lowered from high to medium and thinking always on. Lower effort cuts output tokens far more than it cuts accuracy — which reads as a downgrade. Pin effort explicitly in your API calls or Claude Code config before you conclude the model got worse.
Update: Fable 5.2 takes the Artificial Analysis crown
Fable 5.2 posts the highest Artificial Analysis Intelligence Index score yet (~65), topping every rival — but independent write-ups put it at roughly double the price for a low-single-digit gain. Reach for it on frontier-hard work; do the math before swapping it in for Opus 5.5 everywhere.
Report: Gemini's human-review flap flares up again
A fresh round of reports says Gemini can funnel linked Gmail and Drive content into Google's human-review pipeline; Google denies feeding Gmail to Gemini. Either way, open Gemini Apps Activity — turning human review off reportedly caps chat retention at 72 hours, which quietly breaks long-running conversations.