Claude Fable 5.2

I got early accees so and ran the same things I run on every new model. Skipping the benchmarks. This is what I noticed and what's worth your time. What's different Speed. First thing I noticed. Anthropic say output is up to 30% faster than 5.1 and that matches how it felt. There's also a fast mode in Claude Code and the API that goes up to 2.5x. Check what it costs before you leave it on. Usage. I stayed on medium effort all evening and never felt I needed high. Anthropic say it uses fewer tokens than 5.1 on hard tasks at medium, so start there. Blocks. Around 55% fewer than 5.1. I don't do bio or security work so I can't test this one. If you kept getting bumped to Opus, rerun the exact prompt that got blocked and tell us. Price. API is $8 in / $40 out per million tokens, down from $10 / $50. Model string is claude-fable-5-2. Free resets. Paid plans that include Fable get one to three usage resets this week, depending on how long you've subscribed. Check your usage page and save your biggest job for this week. Vision. Listed as one of the main upgrades, next to coding and reasoning. I gave it a screenshot of a page that looked wrong and asked what was broken. It worked the problem out from the image alone. Fable 5.2 is out. Here's what's actually different and what to try first (cheaper, faster, way fewer blocks) Got on it as soon as it showed up in the model picker and ran the same things I run on every new model. Skipping the benchmarks. This is what I noticed and what's worth your time. What's the same Same model as Mythos with the cyber and bio safeguards on top. Mythos is only for verified security and life sciences orgs. Still won't know what changed on a webpage last week unless you tell it to search. Still eats limits faster than Opus or Sonnet. Use it for the hard thing and the cheaper models for the rest. What to try first Get it to review whatever your last model built. I pointed it at a project my last model built and asked for bugs first, opinions second. It came back with things the old one had walked straight past. Still the best first test for any new model. Give it a screenshot of a broken UI or a photo of a whiteboard and ask what's wrong. Hand it a half-formed request and ask what it would do and why before it does it. I did this on purpose. The plan made sense, which is more than I can say for my request. Rerun the one task 5.1 kept getting wrong, same prompt, so you can compare fairly. May be dont try yet : A giant refactor on day one. High effort for everything. Using it as a search engine. If you find something it does that 5.1 couldn't, post it below. Right, back to it.