Confluence for 6.14.26
72 hours with Claude 5 Fable.
Welcome to Confluence. Earlier this week, we published a piece written entirely by Claude 5 Fable, the first publicly available model in Anthropic’s new Mythos class of models. We’re glad we published that when we did, because by Friday night, Anthropic had revoked general access to Fable due to a government order. This feels, to us, like a turning point in the trajectory of frontier AI. So this week’s edition will focus entirely on Fable: its capabilities, our experience with it, the events of Friday evening, and what we think it all means.
72 Hours With Claude 5 Fable
A release, a revocation, and an uncertain path forward at AI’s frontier.
The Release
On Tuesday, June 9, Anthropic released Claude 5 Fable, “a Mythos-class model” whose capabilities exceeded anything Anthropic (or any other lab) had released to date. Mythos refers to the model that Anthropic announced in April, which the company deemed too powerful to release to the public and instead released to a small set of organizations for cyber-defense testing in an initiative dubbed Project Glasswing. Fable was Mythos-class, with added safety measures for general use.
Benchmark performance data suggested Fable was as powerful as Anthropic claimed. The Anthropic release cites a range of benchmarks, with Fable leading on all but two (which were led by Anthropic’s Mythos Preview model). Commentators with early access agreed. Ethan Mollick wrote that Fable “represents a very real leap over every model I have used before, and, maybe more important, suggests our relationship with AI is changing in drastic ways.” Zvi Mowshowitz declared simply that “Claude Fable 5 is the new best publicly available model” before elaborating: “I have noticed a step change, where Fable can suddenly help me in ways that previous models were not worth bothering to query. Almost everything it has noticed in one of my drafts so far has been spot on and it is downright scary.” X and Substack featured dozens of posts from AI insiders showing Fable accomplishing things that previous models couldn’t.
But the release was not without controversy. Close readers of the model’s 319-page system card quickly spotted language revealing that the model would silently degrade its own responses on certain frontier-AI-development tasks, using invisible techniques like prompt modification rather than an honest refusal, and without telling the user it was doing so. In other cases that Fable deemed risky or dangerous, it would downgrade the user’s chat to a lesser model, Opus 4.8. It didn’t take long for X to light up with early users sharing their frustration with Fable’s refusals. Andrej Karpathy, who recently joined Anthropic, praised the release while observing that the safeguards were tuned to be “a little too trigger-happy for launch.”
A second factor complicated the rollout: while Fable was immediately available in most paid Claude plans, Anthropic noted that on June 22 the pricing model would shift. Rather than Fable being included in subscription plans, it would move to a pay-per-use model. We’ve been expecting a shift like this and shared some thoughts at the time of the Mythos announcement. We expect this shift will go beyond the Claude models, and as it unfolds, it will have major implications for organizations and the decisions they need to make about AI use.
Suffice to say, it was a rocky rollout that was in many ways unprecedented. Anthropic was wrestling with the tradeoffs that come with releasing such a powerful model and the new guardrails and financial considerations it required. But we were excited to try Fable. For the next three days, we did.
Our Experience
Our initial experience with Fable was consistent with Mollick’s and Mowshowitz’s. Fable could simply do more than previous models could, and it could do all of those things better. As our Confluence team discussed our initial impressions, we all agreed that while it was hard to articulate or pinpoint precisely, Fable seemed to “get it” in a way that other models didn’t. It could quickly pick up on context and nuance. It didn’t require much hand-holding: just a goal and the same type of context we’d give a human to whom we were delegating real work. Its first attempts were always close to the mark, and often were nearly perfect.
An example of what this looked like in practice: One Confluence writer used Opus 4.8 to create a suite of files to use as context for a client demonstration. The files, all related to a fictional factory closure, included AI-generated strategy memos, operational plans, factory analysis, and more. Opus 4.8 got the job done. When Fable came out just a few days later, this writer asked it to make the entire suite more realistic. Rather than adding detail or files, Fable added a decidedly human-feeling texture. It changed file names so they followed multiple naming conventions, mimicking the sense of working in a shared folder with multiple contributors. It added contradictions in the facts, as happens all too often when working with a large team. It converted memos to PowerPoint files and tweaked fonts and formatting throughout to make the whole set feel messier, like the file set was built over time rather than in one fell swoop. We didn’t prompt Fable to make these specific changes. We just gave it a goal, to make the files feel more realistic, and it got to work.
Ethan Mollick captured this experience well in his post. Working with the model, he wrote, was less like chatting with an assistant and more like commissioning a small studio to take on a project and return with it finished. This is a step change from the “naïve intern” analogy that many have used for so long. Fable certainly is not an intern, and it does not feel particularly naïve. Mollick notes that working with a model like this is another shift from “casting spells” (carefully crafting each prompt and steering every step) to commissioning work (describing what we want and waiting for it to be done). We agree.
On Friday afternoon, two Confluence authors had a short text exchange about a range of interesting things we wanted to try with Fable, given the new possibilities. A few hours later, our team had a text exchange that was entirely different.
The Pullback and What It Means
On the evening of Friday, June 12, Anthropic posted that it was suspending access to Fable and Mythos, effective immediately:
The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees. The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance.
Within an hour or so, the model was no longer available, and at the time of this writing that is still the case. We’re writing this on Saturday, and speculation continues to swirl about the events and conversations that led to this decision. The story as it stands at the moment is that engineers at a U.S. government partner (the speculation is that it was Amazon) identified security vulnerabilities in Fable that could cause users to avoid certain of its safeguards. There are varying stories about what happened next, and whether the U.S. government asked Anthropic to take reasonable or unreasonable steps. You can read most of the sides of the story in a nice roundup here by Zvi Mowshowitz.
Regardless, right now, no one knows when Fable will become available again and, presuming it does, what safeguards or restrictions Anthropic will add to satisfy the U.S. government’s concerns. No matter what comes to light or what happens with Fable, we’ve entered uncharted territory.
Since OpenAI released ChatGPT in November 2022, it’s been easy for the public to access near-frontier AI models. The AI labs have been quick to release models in part due to competitive pressures but also from a stated belief that it’s better to get the most powerful models in the hands of the masses so we can better understand the technology’s capabilities. Any person willing to spend $20 per month could access models only a few months behind what the AI labs themselves were using. That is no longer the case. Fable and Mythos have been deemed powerful enough (or risky enough in their current state) that the U.S. government has limited access to them.
It’s not unusual for governments to regulate access to technologies that present a reasonable risk. We don’t allow just anyone to access nuclear materials or biological agents. But there also isn’t widespread demand for those technologies. Fable and Mythos are different. There is widespread demand in the market for their capabilities. They have broad applications beyond the risks they present. And notably, they are a general-purpose technology. We simply don’t have a precedent for what the regulation of a general-purpose technology looks like, and it is difficult to think of a government restricting access to or development of previous general-purpose technologies like electricity, steam engines, or the transistor.
That all said, the toothpaste is out of the tube. Our personal reaction to all of this was, oddly enough, a sense of loss. For three days we were excitedly doing things with Fable that we could not do before, and our aperture on what it meant for our work within the firm and for clients was quickly opening … until it shut. As amazing as Claude Opus 4.8 is, we feel like we’ve experienced a regression, and we’d expect others may feel the same. We eagerly await Fable’s return.
Until then, we don’t know who will have access to Fable, Mythos, and future, even more capable, models. We don’t know what that access will look like. We don’t know if we can trust those models to stay available once released. We don’t know what they will cost or who will be able to afford them. We don’t know what this means for how generative AI proliferates through organizations and society at large. What we do know is that uncertainty is generally worse than certainty, and that we’ve crossed a threshold in how governments and AI labs approach access to their leading models. And it’s hard to imagine going back.
We’ll leave you with something cool: Six hours before Fable was revoked, Anthropic posted an X thread of ambitious projects people had already built with it.
AI Disclosure: We used generative AI in creating imagery for this post. We also used it selectively as a creator and summarizer of content and as an editor and proofreader.
So, will OSS keep playing catch-up, or will the frontier labs trip and let it pull ahead?