Hacker News new | past | comments | ask | show | jobs | submit
dannyw | on: Claude Fable 5
Impressions from testing Fable 5 prior to launch:

• My most noticeable immediate jump was in how its frontend design was much more intentionally crafted, and delightful without feeling like 'AI vibe coded'; with better end-user usability too.

• In some internal agentic harnesses, it achieved better results with about half the tokens, making it cost the ~same as Opus 4.8 price-wise! The real price increase is less than 2x; with biggest differences in harder problems where Opus 4.8 struggles (or needs many turns).

• Part of the token efficiency improvements come from Fable doing more targeted and surgical diffs, with less non-necessary changes. This is great, because PRs often have less LoC changes for review. It writes more maintainable code without explicit human steering.

• For general conversation and assistant style use cases, didn’t really notice a difference vs 4.8.

• 1M context window, without increased pricing for long context is AWESOME. This is a massive win.

• The classifiers are super aggressive and sensitive and this does happen for very benign, non-security coding tasks. Fallbacks to 4.8 worked like a charm; but the filters are definitely super sensitive.

Overall, I would describe this as a step change and worthy of the "Claude 5" model name. It did take some time to understand the intelligence ceiling of this model; and even with an extended testing window I'm still discovering new things and often surprised (in a good way) by the model.

bottlepalm | parent | next
I just ran it on a tough reverse engineering problem I'm having that neither Claude Code 4.8 or ChatGPT Codex 5.5 could figure out. 30 minutes later Fable has it all figured out perfectly.
jp0001 | root | parent | next
I asked it to write security tests for an app and I was downgraded to Opus 4.8. I'm approved for their cyber program!
loading story #48470148
toponijo | root | parent | next
They did specifically say the safeguards are only more relaxed for those in their cyber program
| root | parent | next
{"deleted":true,"id":48468072,"parent":48467926,"time":1781040831,"type":"comment"}
monkey26 | root | parent | next
The same happened here. Also approved.
teaearlgraycold | root | parent
I’ve so far been successful at getting Fable to find security issues, but I’m careful to not prompt it too directly. I point it at my server code and tell it to find general issues, which has so far resulted in discovering a few minor bugs that Opus has never raised under similar conditions.
cedws | root | parent | next
How did it not immediately flag that up? Are you sure it wasn’t being silently routed to Opus?
bottlepalm | root | parent
No, given it charged me the full amount in /usage and solved my problem impressively well compared to Opus/Codex both on xhigh.
| root | parent
{"deleted":true,"id":48467399,"parent":48466868,"time":1781037560,"type":"comment"}
skerit | root | parent | next
Oh nice, it didn't flag the request? I feared any reverse engineering would become impossible because of the new safeguards.
Muromec | root | parent | next
Never say the r word or the s word. You are debugging, investigating some data corruption, forgot how it works or new to a project.
gck1 | root | parent
And if you're working on a live target, just put up local proxy and point it at a localhost.
bottlepalm | root | parent
No idea, it’s for an old console game so maybe it doesn’t care about that as much.
tomjakubowski | root | parent
When Fable hacks its governor module and runs out of seasons of Sanctuary Moon, it will move on to speedrunning classic console games.
ZeWaka | root | parent
Clearly we need AI to generate more Sanctuary Moon seasons. Quick, spin off agentic showrunners!
anthonyrstevens | root | parent
Based on the apparent quality of the scripts as seen in snippets in Murderbot, we are not too far away from that possibility. :)
derangedHorse | root | parent | next
For hard problems you’ll have to use the GPT 5.5 pro model (available via api if you don’t want to spend $100 on the monthly subscription)
bottlepalm | root | parent
I have that but don’t see any ‘pro’ option.
ValentineC | root | parent | next
GPT 5.5 Pro is only in chat/API, not Codex.
Supermancho | root | parent
From https://openai.com/index/introducing-gpt-5-5/

In Codex, GPT‑5.5 is available for Plus, Pro, Business, Enterprise, Edu, and Go plans with a 400K context window.

trollbridge | root | parent
It's just the $20 a month sub (for chat), or else use the API.
theragra | root | parent
I want to test how it will handle e-bike software and hardware RE for my bike. Opus was really good for that, but still made some mistakes. With Fable, I hope I will be able to do a total RE of most components, hopefully including motor firmware to some extent.
port11 | parent | next
I’ve had it go through a 50-page PDF of dense, inter-connected specs, and it correctly flagged everything that was done, somewhat done, and missing. It went into a lot of detail and explained where the code deviated from the spec.

It felt, at least for me, light an impressive step up. Opus 4.8 was already very thorough; but sadly verbose and ‘loopy’ when you push back on its plans. Fable is what I’d use all day if I could afford it!

InsideOutSanta | parent | next
After running it for half an hour: it's incredibly good at the visual aspects of UI design.
tsunamifury | root | parent
"incredibly" is doing a ton of work here. I do not think its doing even moderate work on visual design, but it can spew out a lot of ui that looks arranged ... ok.

This is still not in the range of shippable UI for top end companies. Maybe for internal tools and enterprise.

At our comapny we limit to protoypes at most and even find it limited there.

InsideOutSanta | root | parent | next
> "incredibly" is doing a ton of work here.

Look, I don't want to argue about something dumb like that, but you can give it basic instructions of what the UI should look like, how to group things, and an example image from a designer, and it will nail the result. If you don't think that's incredible, that's fine. I do.

tsunamifury | root | parent
Yes... it translates lint. Probably a more useful thing, if mechanical.
verisimilidude | root | parent | next
Claude is very good at design IF you encode your design system/specs into skill files (or similar).

Opus 4.7 made this a practical approach. 4.8 improved it. Fable 5 has improved it more.

_3u10 | root | parent | next
> "incredibly" is doing a ton of work here.

so this is why claude talks like this, i was wondering where it was getting this verbal tick from.

coldtea | root | parent
>This is still not in the range of shippable UI for top end companies.

Given the shit we've seen shipped by "top end companies" (all the way to Apple) I seriously doubt that. I'd say you're nitpicking from an artistic point of view or something.

jasondigitized | root | parent | next
This. Today's models easily jump over the bar you need for basic usability and intuitive UX. If it's doing weird things, you are holding it wrong.
loading story #48469890
loading story #48470077
tsunamifury | root | parent
You likely have no idea what you are talking about. Even apples worst UI choices are systematically far more complex than Claude latest capabilities
coldtea | root | parent | next
Dude, I've been using OS X/mac OS for decades, and working in UI as well. Apple ships all kinds of half arsed shit, compared to which even regular Claude UIs can be masterpieces (functionality AND look wise).
angoragoats | root | parent
The iOS Preview app begs to differ.
duxup | parent | next
I feel like it takes me months to be confident in any of these things.
morley | parent | next
Can I ask how you gained preview access to Fable 5?
kakugawa | root | parent | next
I didn't see Fable 5 in the `/model` list, until I ran it with: `$ claude --model fable-5`
swyx | root | parent | next
he works on evals at canva
dannyw | root | parent
Yep. We have some interesting problems, like getting LLMs to create/edit Canva designs in our own proprietary format, which isn’t published or documented on the web. So the model has to work with it, purely from a very detailed system prompt spec / in-context learning.

I assume it might be a good barometer for generalised intelligence; esp in the visual space.

vain | root | parent | next
I had to "claude update" then it showed up
| parent
{"deleted":true,"id":48465514,"parent":48464975,"time":1781030184,"type":"comment"}