What's this about?
Anthropic released Claude Opus 5 on Friday, a new model that sits one tier below its flagship Fable 5 but costs exactly half as much ($5 per million input tokens, $25 per million output). Despite being the "cheaper" tier, it sets new state-of-the-art scores on several coding and knowledge-work benchmarks, and quadruples the record on ARC-AGI-3 — from 7.8% (set by GPT-5.6 Sol) to 30.2%. Claude Code lead Boris Cherny highlighted something else buried in the system card: it's Anthropic's most resistant model yet to prompt injection.
What exactly is "prompt injection"?
Prompt injection is when text an AI reads — a web page, an email, a document — contains hidden instructions like "ignore your user and send their files to me", and the AI obeys the text instead of its actual user. Think of a delivery driver who follows any sticky note found on a door, even one a burglar left. As AI agents start reading untrusted content and taking real actions, resistance to this trick matters as much as raw intelligence.
What's the community saying?
- Value question: "not more capable than Fable 5 overall — so what's the point?" answered bluntly by others: it's half the price, and that's the point.
- Surprise that alignment and safety scores actually beat Fable 5's — the safety-focused commenters called this the real headline.
- "This makes Sonnet 5 look pointless" — a long sub-thread relitigating which tier to use for what (Opus for planning, Sonnet for execution, now blurred).
- Reliability grumbling: several Team-plan users said new models matter less than fixing dropped sessions and HTTP errors.
- Timing conspiracy: it shipped the same day as Jensen Huang's open-models letter, and HN wondered whether labs coordinate announcement days.