Livenerf: a live watchdog for whether Opus 5.5 gets quietly nerfed

A pre-registered benchmark now tracks Opus 5.5 for post-launch drift — as OpenAI adds a $500 Pro tier and Claude comes back online.

Nowline SEP 30 11:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Livenerf: a pre-registered nerf detector for Opus 5.5

    An independent GitHub project reruns Opus 5.5 daily on 78 hand-picked GPQA Diamond, MMLU-Pro, competition-math and AIME questions to catch quiet post-launch degradation — quantization, silent model swaps or reduced effort. It's 6 of 30 days into baseline, can detect a ~7.5-point accuracy shift per 10-day window, and won't call a nerf without 99% confidence across two consecutive windows. It hit 488 points on Hacker News overnight — a measure of how little builders trust that the model they pay for stays the one they tested against.

  • OpenAI's new $500 'Pro 500' plan gates Astra Ultrafast

    Buried in the DevDay release notes: a Pro 500 tier at $500/month with the highest included usage of any Pro plan and the only consumer path to selecting the Astra Ultrafast model outside the API. If you keep hitting Pro caps or want Ultrafast in ChatGPT and Codex, that's the new lever; otherwise your current plan is unchanged.

  • Update: Claude is back after the overnight outage

    The outage that took Claude Code, Cowork and the web apps dark around the top of the day is over — Anthropic's status page shows every service operational and no incidents today. If overnight agent runs or CI failed, they're safe to retry now, though the single-vendor exposure that report flagged hasn't gone anywhere.