Fable 5 is back on Claude.
Unfortunately Claude is now broken on all my browsers for paste of larger text blocks. I have to put them inside files and drop those in. What a PITA.
Fable 5 is back on Claude
I have tried it during the first window nd again since it came back this week. Cant say I “feel” a huge difference from opus for what I do except my token window expired noticeably quicker. I also kick off a few reviews of security, meteorology, efficiency of my app in parallel opus/fable and got a few issues implemented in two separate PR from both opus/fable. Not overwhelmed by the difference.
I feel fable maybe worth only if you do really advanced stuff… right now I don’t intend to upgrade my plan to use more of fable…
By the way I had done the similar experience (same issue two models implementation codex vs opus) and in that case it was clearly different (opus won, even codex when evaluating both thought opus was better!)
I find it dramatically better if the task is complex and larger in scope. And you need to be giving it a goal rather than instructions to get the most out of it.
Also (noting that they keep changing on this, but as of right now I believe) it’s not going to be useable on subscription after next week, only on paid API usage, which is quite expensive for hobby/experimental use. Easy to spend £100 on a single day’s session.
Do you mean Claude will no longer be available on a subscription?
That would be incredible. It would vanish from most common usage.
Or is this just Fable 5.
Fable 5 will not be available on subscription. Nothing changing with the other models as far as I am aware.
I did one task with Fable 5 (albeit fairly involved, but it was a single “ask”) and it came to $230.
David wrote:
And you need to be giving it a goal rather than instructions
Interesting. I need to figure out how to integrate that (goal) into my workflow. Though a bit scared of the cost increase potential… My tasks are somewhat big, they can take a few hours of agent time, but I usually upfront it with a few big brainstorm/planning sessions that leave a few issues artifacts that then parallel agent go an executes in parallel. But I don’t really see yet how I can replace the brainstorm session with a “goal” because often it’s not well defined, I like to discuss and explore different options etc. So with a “goal” you let the ai decide everything? I’d love to hear about some concrete examples of a big tasks run with a goal autonomously.
Some example for me (were I have fable vs opus or opus vs codex) would be things like complete refactor of the iOS app including upgrading all the cross section layers, skew t etc, or add a new framework to generate time base scenarios to find better time to do a flight. These required thinking through different ux, ui options, how to make the pipeline asynchronous and different steps to feed back update to the UI etc. I don’t know how I would have defined these as a goal given I was not sure how it should look like at the end
It’s going to be interesting when they do their IPO. Sales about 10BN, flotation size about 1TN 
Makes the Dutch tulip bubble look trivial.
Google is 4.3TN but Google gets money from just about everybody who is in business. Even a tiny business is paying them a few hundred per month for adwords.
So with a “goal” you let the ai decide everything?
No. In fact, it sounds like you’re essentially already doing it right. Have lots of back and forth discussion on what you want to build, which is plenty good enough with Opus. Have it grill you, do design work, examine the existing code base, etc. etc. Ensure that it includes acceptance criteria, tests, etc. Get it to generate a comprehensive PRD, referencing any other documents that it generates or that already exist. Then clear your context, switch to Fable, plan mode, point it at the PRD and tell it to build it without interruption and don’t stop until it’s done and meets all the requirements as stated. It’ll verify the PRD, draw up a plan and validate it against your existing code base (which might be overkill, given it has only just done that as part of drawing up the PRD, but I think it’s worthwhile as a second chance to catch anything). Then tell it to go. The goal is implementing that PRD. Very important: give it access to a browser which can exercise what it’s building.
From here
If you can get your hands on the software, you can find the key. Unless it does something like using Diffie-Hellman to generate new keys, but in a car this seems unlikely.
However I have a cautionary tale. Claude offered to crack a binary file format for me, which involved reverse engineering the encryption. It did all this without me ever saying more than “yes”. But now Claude has cancelled my account for “policy violation”, which I presume means this. (They refuse to tell you what policy they think you have violated).
No great loss. But if you want to crack the crypto and can get access to the binary file, it’s not hard but do it from a burner account!
