Fable 5 is the real deal! Managed to find a good solution to a viewport/device bug on out website that none of the previous generation models ever found a good solution to.
Congratulations Anthropic!
Fable 5 is now available in Claude Code and Cowork
Fable is the best model I have used for coding, by a wide margin. It is a big step up, enabling less prompts and steers, more efficient token use, better code quality, better tool use, more intelligent self-verification, longer running sessions, and higher trust & autonomy.
Happy coding!
Introducing Claude Fable 5: a Mythos-class model that we’ve made safe for general use.
Its capabilities exceed those of any model we’ve ever made generally available.
@asaio87 The trend is autonomous development. I think we end up there, but honoestly, it's a bit further away than current capabilities of harnesses/code bases.
The AI moat is shifting from model engineering to context engineering.
The best agents don't win because they reason better.
They win because they see the right information at the right time.
@leerob Coding agents work so well you immediately want to run 5 of them in parallel. But in my experience that is when the real chaos starts (and what we are trying to solve at blume :) ).
Engramme challenges the current search based memory paradigm. And proposes human-like recall memory.
Very interesting paper with potentially profound impact on AI, but sceptical of the romanticization of human memory
Transformers are stateless. The world needs a post-transformation era to solve the problem of memory. We are making searching and prompting obsolete with Large Memory Models. View our paper here: https://t.co/2MUkDKu3rT
The coding agent example in section 3 is exactly what we keep running into - agents that rediscover the same codebase idioms every session.
What you're describing at the architecture level, we're trying to patch at the intent layer: externalizing decisions and patterns so agents don't have to re-infer them. Band-aid vs. root cause, probably.
@quxiaoyin The limit isn't cognitive - it's tooling. When you have a shared source of truth for intent, you stop being the context router between agents.
@Pragmatic_Eng@thdxr The landmines form because agents don't share a living model of intent. Each session starts fresh and ready do put in new landmines, not remembering previous intents.
Opus 4.8 (high) really likes burning tokens.
Just watched it spend 30 minutes chasing a corner-case PostHog bug. It kept trying to test the same failure mode through different approaches, failed repeatedly, and never gave up.
Impressed by the effort, not by the token spend.
@OpenAI Just back from a 5 hour sesh arguing with Codex on non-flicker posthog setups. The day agents works with complex sites I will be overjoyed! But not there yet
Jensen Huang just shared one of the craziest charts in software history
- Commits: 1.4B
- Pull requests merged: 90M
- New repositories: 20M per month
3x more commits and pull requests in 2026 than all of 2025
Useful AI is here