Anthropic's Fable a cautionary tale

Presented by:

Welcome to Runtime! Today: Anthropic's first Mythos-class AI model launch did not sit well with a lot of enterprise users, Oracle investors remain wary about its expansion plans, and the latest enterprise moves.

Please forward this email to a friend or colleague! If it was forwarded to you, sign up here to get Runtime each week.

The scorpion and the frog

After the launch of AI models eight months ago that finally delivered on the promise of agents, Anthropic seemed poised to capture the lion's share of enterprise workloads. But following weeks of hype about the too-dangerous-for-prime-time capabilities of its Mythos-class models, the first one out the door fell flat on its face.

Anthropic released Claude Fable 5 and Claude Mythos 5 Tuesday as its latest topline AI models, with a special emphasis on their ability to handle long-running tasks. "It is state-of-the-art on nearly all tested benchmarks of AI capability, showing exceptional performance in software engineering, knowledge work, vision, scientific research, and many other areas," the company said in a blog post.

Both models outperform the Mythos Preview model released in April as part of Project Glasswing, but Fable 5 is a locked-down version of Mythos 5 available to the public with safeguards that reroute prompts for questions around cybersecurity, biology, and model distillation to Opus 4.8, the previous flagship model.
Mythos 5 shipped without those safeguards to Project Glasswing participants, and Anthropic said it planned to release it to a wider group through some sort of "trusted access program" that appears distinct from Project Glasswing.
Both models will cost $10 per million input tokens and $50 per million output tokens, or "less than half the price of Claude Mythos Preview," Anthropic said, but that's also twice the price of Opus 4.8.

It was clear that Anthropic knew the restrictions placed on Claude Fable 5 would be controversial: "... we’ve deliberately tuned the safeguards to be cautious, and they are still stricter than would be ideal," it said in the blog post. But the heavy-handedness with how those safeguards were applied took many users aback.

Matt Suiche told Techcrunch that Fable 5 balked at simple tasks like code review: "if you ask it to write secure code, it assumes it is cybersecurity related work instead of software engineering best practices, and you get downgraded," he said.
Business Insider asked Fable 5 some very basic questions about cancer, and it flagged those questions as inappropriate for the model with a pop-up window.
However, a detail that didn't appear in the main blog post drew the most ire Tuesday and Wednesday: Fable 5 was designed with "new interventions that limit Claude’s effectiveness for requests targeting frontier LLM development (for example, on building pretraining pipelines, distributed training infrastructure, or ML accelerator design," Anthropic said in the very dense system card released along with the model.
That means unlike the pop-up Business Insider encountered, when "Claude gives me poor or incorrect advice while I’m working on an AI component, I have no way of knowing whether the model was confused, whether my problem is unsolvable, or if some invisible policy restriction quietly kicked in. Anthropic has explicitly chosen not to tell users when this is happening," wrote Jonathan Ready, founder of Wanderfugl, on his blog.

That decision did not engender a lot of trust among those working on the cutting edge of enterprise AI, and Anthropic's decision to start retaining corporate data fed into Claude Fable 5 may mark a turning point in its relationship with those users. In a substantial departure from its previous policies around handling enterprise data, "prompts submitted to, and outputs generated by, Mythos-class models [will be] retained for 30 days for trust and safety purposes, on every platform where these models are offered," including the Big Three hyperscalers, Anthropic said.

Anthropic said it would institute strict policies around how internal employees handle that data, but it doesn't really matter; data retained is data that can leak.
Microsoft immediately restricted access to Fable 5 for internal employees based on the new requirements, according to The Verge.
Anthropic cited the safety concerns around its Mythos-class models as the reason for the policy change, but that only raises the question of why the company felt compelled to release the model to enterprise customers when it is so apprehensive about how the model might be used.
In the meantime, check out Aesop's "The Fox and the Goat."

A MESSAGE FROM GITLAB

GitLab Orbit gives your AI agents the context to move at full speed. Smarter agents, faster results.

Learn more at GitLab.com

Heavy lifting

Oracle's attempt to capitalize on the AI boom is by far the most precarious of the old guard IT companies, forcing the company to burn enormous amounts of cash in hopes of building an infrastructure service that can rival the Big Three. Any eventual payoff will be years in the making and will require even more investment than previously thought, Oracle told investors Wednesday.

The good news is the company beat Wall Street's expectations for fiscal fourth-quarter revenue and profit, and it raised its profit guidance for the current quarter, according to CNBC. The bad news is that management told investors it will need to raise an additional $20 billion to meet its infrastructure goals, and those investors sent its stock down nearly 10% in after-hours trading.

Oracle's fate is intertwined with OpenAI's ability to generate enormous revenue over the next several years, and even if everything goes perfectly that will still be a difficult task. "Oracle has a lot on its plate. They have a CEO [CTO, technically] who is very distracted with his other side quests in the media business," Cleo Capital's Sarah Kunst told Bloomberg.

Enterprise moves

Paul Colangelo is the new CEO of Casepoint, joining the government-focused enterprise software company after serving as founder and CEO of Neumo.

Joel Molinoff is the new chief operating officer of iCOUNTER, joining the threat-intelligence company after similar leadership roles at BlueVoyant and CBS.

John McCauley is the new chief financial officer at Vanta, joining the compliance software company after finance leadership roles at Calendly and Seismic.

Casey Watson, Jason McClelland, and Sherri Manning are the new chief revenue officer, chief marketing officer, and chief human resources officer, respectively, at Sonatype.

The Runtime roundup

GitHub announced several changes to the way the next version of npm will handle package updates, requiring explicit approval before installing new packages in hopes of tamping down the wave of supply chain attacks on package managers.

ServiceNow informed customers that their data might have been exposed through a security flaw that security researchers were able to identify and has since been patched.

Salesforce announced plans to acquire m3ter, which helps companies roll out consumption-based pricing schemes as it transitions away from per-user pricing.

Google Cloud customers in India were impacted by a fire at a third-party data center in Delhi, affecting networking workloads across the country.