lucas g

teaching models to use computers at anthropic. prev: google deepmind, google brain. opinions are my own.

[Claude @claudeai]

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.

[lucas g @DaLucasGonzalez]

@hopes_revenge chill with that please

[lucas g @DaLucasGonzalez]

opus 4.8 is out, i find it to be an incredibly pleasant and capable model to work with. we also wanted to drop more goodies to address your feedback:
- we increased rate limits on claude code
- same price on opus while more token efficient
- significantly cheaper fast mode
- from the blog: "we expect to be able to bring Mythos-class models to all our customers in the coming weeks"

[lucas g @DaLucasGonzalez]

@MeganNyvold 210c for 15p is a bit tough not sure i’d classify as a protein bar really..

[lucas g @DaLucasGonzalez]

@SynabunAI @LucaWeihs send me failures and we’ll improve , dms open

[lucas g @DaLucasGonzalez]

easter egg: we paired this with an open source demo you can play with or immediately build a UI on top of, linked in thread. claude code will walk you through the whole setup! was fun to work on this with another Luca(s) @LucaWeihs

[ClaudeDevs @ClaudeDevs]

Computer use turns Claude into an agent that can operate real UIs.

New blog post on making it reliable in production: getting click accuracy right, choosing thinking effort levels, keeping long sessions within context, and recording demonstrations Claude can replay:

https://t.co/JwSI3YDWcK

[lucas g @DaLucasGonzalez]

@bil0090 @ClaudeDevs did you try some of the patterns for improving latency? haiku + advisor model or with saved learn mode examples is very quick

[lucas g @DaLucasGonzalez]

@renefaurskov @ClaudeDevs i suspect this will be table stakes in the near future, analogous to a business having a website today

[lucas g @DaLucasGonzalez]

@SynabunAI @ClaudeDevs if the surface never changes, selenium works. This is a way to get a cheap CU model (e.g. haiku 4.5) to have greater realiability and auto heal when things change

[lucas g @DaLucasGonzalez]

@chimpansky @ClaudeDevs check out the "Recommendations for effort levels" section, we included effort curve graphs

[lucas g @DaLucasGonzalez]

@zzxwill @ClaudeDevs check out the linked demo: https://t.co/Xfc4qXLIm3

we also include a tool where you can test clicking specifically on a static image

[lucas g @DaLucasGonzalez]

@LucaWeihs https://t.co/Xfc4qXLIm3

[lucas g @DaLucasGonzalez]

@henrytdowling @RLanceMartin use it if you are price insensitive & doing a CU task that also involves complex reasoning - e.g. like using professional software to complete a difficult task

[lucas g @DaLucasGonzalez]

@tszzl hope my side gig gets in

[Adam Wren @aswren]

Dawkins is more intelligent than 99% of the people making fun of him and ‘if AI can be just as capable as us without being conscious, why did we develop consciousness in the first place?’ is a great question

[lucas g @DaLucasGonzalez]

we’re basically on turn 400 of a game of civ and the US is yoloing a science victory which requires datacenter buildings in each of your cities and a few great scientists and technologists

[lucas g @DaLucasGonzalez]

@Ggod323 compaction

[lucas g @DaLucasGonzalez]

@magnushambleton @tenobrus literally

[lucas g @DaLucasGonzalez]

@DanielleFong @theo what are some of the rough edges you're seeing?

[Andon Labs @andonlabs]

Claude Opus 4.7 is pretty good at Vending-Bench 2