arxiv.org/abs/2501.00536…

Really excited to share what we’ve been working on with an amazing team! I previously thought adversarial robustness was completely doomed. Robustifying a single LLM policy is really hard, but with multiple overlapping monitoring systems, we’re seeing promising results.

Introducing Claude 3.7 Sonnet: our most intelligent model to date. It's a hybrid reasoning model, producing near-instant responses or extended, step-by-step thinking. One model, two ways to think. We’re also releasing an agentic coding tool: Claude Code.

"why isn't claude 3.7 sonnet better at esoteric competition math problems" we found it didn't generalize to becoming a pokemon master

I mean, surely you saw this coming 😛

this is an extremely accurate simulation of what it was like to work on 3.7 sonnet. we are hiring

`npm install -g @anthropic-ai/claude-code` there's no more waitlist. have fun!

my codebase after vibe coding with Claude for 6 hours

another Eval That Actually Matters

We've made Claude Opus 4 and Claude Sonnet 4 significantly better at avoiding reward hacking behaviors (like hard-coding and special-casing in code settings) that we frequently saw in Claude Sonnet 3.7.

THE WAY OF CODE, a project by @rickrubin in collaboration with Anthropic:

Claude decided it was time to go to bed so it drew some goodnight ascii art and ran time.sleep(28800). I’m dying.

Better than opus 4.1 at sonnet prices - enjoy!!