Fable 5 will default to Opus 4.8 for coding tasks | Hacker News Reader

Fable 5 will default to Opus 4.8 for coding tasks

xcancel.com

49 pointsbabelfish30 comments

If fable 5 can't do code... what are people using it for?

Was wondering the same. So coding, debugging and anything that has to do with cybersecurity or LLM development is blocked. What tasks are left in which Fable 5 is meaningfully better than the predecessors?

It's beginning to look like giant publicity stunts. A model that can't do what it says it can, overhyped, no way to test.

It's good for pre-IPO PR.

You drop a good old "A new {cool_word}-class model.".

And boom! What are you competitors going to do? Use same classification nomenclature? I don't think so! IPO secured.

How can it be legal to charge a higher fee for the Fable model, if most of the usage will default to a cheaper model behind the scenes?

It is one thing if I, as the user, choose to down-level but Claude shouldn't do this on its own.

They are transparent. Don’t use it if you don’t want to. These things shouldn’t be illegal.

I agree that it shouldn't be outright illegal, but I will definitely be making the personal choice not to engage with model providers which do this. There's no amount of impressive results that could make me want to secretly pay more for less.

Originally at least the switch wasn't silent and whether to halt or auto switch was a setting in Claude Code.

because it's all indistinguishable from magic when it's a cloud SaaS.

Not like you can tell the difference if you dont own any of the implementation.

Judging from their site, I think it should be OK:

"You won't be charged Fable prices for rerouted requests. Learn more about how the fallback experience works."

https://www.anthropic.com/claude/fable Under "Safeguards"

And what happens if you also get blocked by Opus?

I couldn't find if they refund the tokens for that session or the customer is SOL.

You're paying in full for the "guardrails" embedded in the system prompt, prompt injections, refusals, fallbacks and everything else that may be caused by the service provider.

This is the new normal.

This has no mention of what happens to the prompt cache, including the "learn more" link.

Knowing Anthropic, it wouldn't surprise me if it will result in a full cache miss/rewrite at fallback, with potentially up to 1M tokens in the context window.

https://platform.claude.com/docs/en/build-with-claude/fallba...

Fable requests that fall back to Opus do not charge the cache miss cost of Opus.

Using only Fable, if I do /usage it will show both Fable and Opus usage, so I don't think it's charging Fable prices for Opus code gen.

I must admit I was drawn to Claude because of their (obviously successful with me) publicity stunt about "being too dangerous to make widely available", when Mythos came out.

It's a shame because I was really looking forward to use it specifically to find potential security holes in my own software.

That being said, Opus 4.7/4.8 have been quite useful already, especially for finding things in the harder to test, non-happy paths.

If I recall, Fable 5 is supposed to be basically Mythos which falls back to Opus 4.8 when dealing with cybersecurity. I wonder if that also includes "finding bugs that could lead to security exploits".

Yeah, I was very excited for Fable to come back to use it for work after using Opus 4.8, but now I guess I'm just excited for Sol/Terra/Luna (unless they have the same restrictions)

It's very likely that OAI models will have even more restrictions. Firstly because now they know what feds will do if you don't tune the safety classifiers towards more false positives and secondly, OAI models were always more restrictive than ANT.

Is Opus 4.8 equally or more capabale for the actual mechanics of the coding work?

It seems to me that having a powerful Fable layer for the planning, coordination, orchestration-type work and delegating to a suitable model for the actual execution of "coding tasks" is perfectly appropriate, if that is the case.

This is a misinterpretation. Fable 5's acceptable use policy has false positives during some coding tasks, and that's what they were talking about. But I've been using it for web dev tasks since it relaunched today, and it's worked fine without fallback.

(On a firmware-customization project involving a ghidra MCP, it triggered and switched to Opus; that was sort of expected.)

This third-party mirror won't load for me.

That aside, they said "some routine tasks like coding and debugging". Since I'm using it for coding right now and it's notably better than Opus 4.8, I think what they mean is that some coding and debugging tasks will fall back.

How's it different from Opus in your experience?

Fable is a lot more confident making decisions for me. Opus tends to create hacky workarounds rather than thinking through the architecture. The difference is so great that when Fable was withdrawn last month I just stopped using Claude, because I did not want to go back to Opus.

One of my first experiences with Fable was giving it an Opus task and watching it work for more than twice the typical time that Opus works and complete the task to a much higher standard of quality than Opus ever did. That was enough to solidify how much better it is already, but I just keep seeing the difference in everything I do with it.

Fable is useless, then. What's the point of re-enabling it exactly?

[dupe] Discussion on this and the rest of the source post: https://news.ycombinator.com/item?id=48740771

Can't wait for China to pull ahead so we have an end to this bullshit

This is a testable claim.

Gave it a small coding challenge. It wrote the code just fine.

crikey just use the real link: https://x.com/AnthropicAI/status/2072163884430229756

No thanks

Now that's what I call a scam.

Reply on news.ycombinator.com