Follow ZDNET: Add us as a preferred source on Google.
ZDNET’s key takeaways
- Fable 5 promises power, but reliability is a concern.
- Sudden downgrades could derail serious AI projects.
- Opus 4.8 is already good enough for my daily work.
The name “Fable 5” is appropriate because, boy, does it have a story to tell. It’s the story of ups and downs, and possibly ups again. It was introduced with great fanfare as the defanged Mythos, a large language model capable of incredible feats but safe enough for nongovernment work.
Also: Anthropic’s new Claude Fable 5 is the same base model as Mythos but with guardrails attached
But then Amazon reported that Fable couldn’t be let off its leash. Amazon purportedly claimed it could do some of the same dangerous cybersecurity exploit detection that Mythos was famous for, but without Mythos’ limited distribution. The fable is that Amazon tattled to the government about this, and the US government responded by slapping export restrictions on Fable.
That was not only to prevent it from being exported outside the perimeter of the US, but also to ensure that no foreign national would be permitted to use it, even if that person was, in fact, physically within our borders.
Anthropic reacted by shutting down access to Fable and Mythos completely to everyone, regardless of what was being done with it, who was using it, or where it was being used. Fable was available for a short time and then went black. It’s still not entirely clear whether Anthropic’s reaction was out of spite or an abundance of caution.
Now, Fable is back. It’s being advertised in its app, in the Claude Code terminal, in the Claude app, and pretty much everywhere. Usage limits have been raised for just a few days, until July 7, to allow people to get a taste of the once-terrible beast that has been chained at the order of America’s security leadership.
But I’m not entirely sure the juice is worth the squeeze, to use a phrase that has been overused and yet is incredibly tempting because it’s so appropriate for this story.
As I’ve been discussing, the rug was pulled out from under users last month. Fable drops itself down to a less capable AI if it fears that it might be too powerful or might be applied to something too dangerous. I wrote about the uproar over that at the time.
Also: Claude Fable 5 secretly throttled AI researchers, and the internet went wild
Now I’m seeing reports that Fable is throttling performance again, but far more often. Reports are telling us that any time there is even the slightest hint of security misuse (for example, using the word “security” in a prompt), Fable drops down to Opus-level work, as if its new capabilities just vanished in a puff of smoke.
So I’ve decided to give Fable 5 a pass, at least for now.
It’s long been my practice that, when a new operating system comes out, I don’t install it on my main computer for a few months. This lets all the initial bugs and kinks work their way out, allows a few updates to be released, and makes sure that the worst of the launch-day problems have been patched successfully. I think it makes sense to apply this practice to Fable 5.
If it stays stable for two or three months (and I know we’re talking AI time here), and the government or Anthropic doesn’t change the rules, and I need something for which I’m willing to pay double the price of the very capable Opus 4.8, then I might use Fable. Otherwise, it just doesn’t seem all that interesting. Here are five specific reasons why I’m skipping it for now.
1. It’s too volatile
I’ve progressed beyond just using AI to test AI tools. I now use Claude Code, Cowork, and OpenAI’s Codex to accomplish real work.
Last weekend, I used a combination of Cowork and Codex to do a mission-critical mitigation to prevent a spam attack that was rapidly turning into a denial-of-service attack. I worked with the two AIs nonstop, until I had a working protection solution in place.
Also: I had Gemini and Claude write my email replies – but only one sounds like me
This is not the sort of work where you want your AI model to simply stop working. As I described with Codex, I’m on a lower-tier plan. I had to eke out more cycles from it a few times by waiting for the timeout to end. It was deeply frustrating. The difference is that with Codex, all I would have to do is pay more money to keep it going. But if I were relying on Fable, and it unceremoniously shut off, it would simply be gone.
I don’t like it when my tools vanish. So I do my best to only use highly reliable tools. So far, Fable hasn’t earned that label.
2. The guardrails are moving
Based on a Reddit thread, it’s not entirely clear what triggers the downgrade from Fable to Opus. I don’t find that acceptable. Sure, we know that AIs are nondeterministic, and therefore not entirely predictable. But it would be nice to know, fairly generally, when a project will be bounced from the uber-model to a less performant intelligence.
So far, Anthropic has not published definitive guardrail definitions. My guess is it can’t. It’s being highly reactive to the threat of a government ban, so something that might be approved today could well be restricted tomorrow.
3. It’s more expensive
According to Claude, “Right now, through July 7, Fable 5 is included on your Max plan (along with Pro, Team, and select Enterprise) at no extra cost, but capped at 50% of your weekly usage limits.”
However, Claude said, “After July 7, Fable 5 comes off subscription plans entirely. Continued access moves to usage-credit billing at standard API rates — $10 per million input tokens and $50 per million output tokens. In practice, that means it stops being part of what your Max subscription covers and becomes a metered add-on.”
Also: AI Model Release Tracker: Anthropic releases Sonnet 5, plus Fable 5 is back
I’m already paying $100 a month for Claude Max and using Opus 4.8. If Fable is likely to drop back to what I already have on subscription, it just doesn’t seem cost-effective to start using API billing for an unproven product. I’ll wait.
4. It could be slow
This one is a mixed bag. I’m seeing complaints like this one on Reddit reporting that Fable 5 is very slow. The Redditor equated it to running Ollama on a local server (which, I can attest, is too slow).
On the other hand, another user on X makes the case that Fable isn’t slower, but has a different pacing than Opus.
Also: Anthropic rolls out Claude Tag, your new agentic AI coworker in Slack
The thing is, whether it’s slower or paced differently, nobody is reporting that Fable is rocket-fast. I spend a lot of time waiting for Claude Code to do its thing. Doubling or tripling its speed might well be worth spending more, but staying at the same speed doesn’t provide enough justification to jump ship.
5. Opus 5 is probably around the corner
Both Fable 5 and Mythos 5 are back and available to their respective allowed user bases. Sonnet 5 was released on July 1. That just leaves Opus and the baby bear model, Haiku, waiting for upgrades to “5” versions. Anthropic hasn’t told me anything about an Opus 5 release schedule, but you can be pretty sure it will have an announcement real soon.
Also: I tested ChatGPT vs. Claude to see which is better – and if it’s worth switching
The other factor is simple. Opus 4.8 is pretty good. I mean, really pretty good. I’ve been using it actively. It has been doing its job. Yes, today it admitted to making some careless mistakes, but who doesn’t? I’m not convinced that I need something that’s another big expense when the current solution is very ably doing its job.
And they all lived happily ever after
So, am I suggesting you avoid Fable 5? Absolutely not. I don’t know what you use AI for, how challenging your projects are, or what your cloud service budget is. All I can tell you is, for my usage as a solo programmer, columnist, and small-business owner, I don’t think I need it now. But that’s me. You be you.
Also: Switching to Claude? Here’s how to take your ChatGPT memories with you
When Fable is integrated into the Max plan and either replaces Opus or is a next-level tier, I’ll probably give it a try. I haven’t really come up against something that requires the extra thinking smarts that Fable is supposed to be able to offer. That’s just for now, of course. Stay tuned. This fable undoubtedly has a longer story arc. There may be dragons.
Are moving guardrails a dealbreaker for you when choosing an AI model for serious work? Let us know in the comments below.
You can follow my day-to-day project updates on social media. Be sure to subscribe to my weekly update newsletter, and follow me on Twitter/X at @DavidGewirtz, on Facebook at Facebook.com/DavidGewirtz, on Instagram at Instagram.com/DavidGewirtz, on Bluesky at @DavidGewirtz.com, and on YouTube at YouTube.com/DavidGewirtzTV.