Nadella Slams Anthropic’s Fable While Microsoft Builds Its Own AI

Satya Nadella told his own Copilot engineers this week that Anthropic’s flagship Fable model is too locked down to be useful, calling its refusals “editorially controlled” in a private meeting obtained by CNBC. The Microsoft chief executive made the comment about a company his own employer just backed with a $5 billion investment.

The remark landed the same week Microsoft’s own family of homegrown AI models, built specifically to reduce its dependence on outside labs, is rolling into GitHub Copilot, PowerPoint and Azure Foundry. That timing is the part most coverage of Nadella’s comments has skipped past.

Nadella Tells His Own Engineers Fable Doesn’t Make Sense

The comments came Wednesday, in a meeting with the team building Microsoft’s Copilot AI software. According to a copy of his remarks reviewed by CNBC, Nadella singled out Anthropic’s Fable 5 model for refusing requests he considered ordinary.

“If you use Fable, when it refuses for any random thing, it just is like, when was the last time you had a creation tool that was so editorially controlled?” Nadella told the engineers. “It doesn’t make sense.”

Microsoft declined to comment on the remarks when CNBC asked, and Anthropic did not respond to a request for comment. The company Nadella was describing is not a rival Microsoft is trying to unseat. It is a partner Microsoft pays into and profits from at the same time.

What Triggered Fable’s Lockdown

Fable’s current restrictions trace back to a security scare, not a policy whim. Anthropic released Fable 5 and its more tightly controlled sibling, Mythos 5, on June 9. Three days later the U.S. government ordered both models cut off for foreign nationals worldwide, after Amazon researchers reported a way to prompt Fable 5 into identifying software vulnerabilities and, in one case, writing code that showed how one could be exploited, according to Anthropic’s own account of the incident.

Here is how the saga unfolded over five weeks:

1. June 2: Microsoft unveils its own seven-model MAI family at its Build conference, months before the Fable dispute even begins.
2. June 9: Anthropic releases Fable 5 and Mythos 5 publicly for the first time.
3. June 12: The U.S. Commerce Department orders both models cut off for foreign nationals, including Anthropic’s own non-citizen staff, after a reported safeguard bypass.
4. June 26: The government approves restoring Mythos 5 access for roughly 100 trusted U.S. companies and agencies.
5. June 30: Commerce lifts the export controls entirely after Anthropic trains a new safety classifier.
6. July 1: Fable 5 returns globally, with Anthropic warning the tightened safeguards would flag a somewhat higher share of harmless requests.
7. July 15: Nadella tells Copilot engineers the resulting refusals “don’t make sense.”

Anthropic says the new classifier blocks the specific flagged jailbreak technique in over 99% of cases, and that blocked requests are not simply refused. They get rerouted automatically to an older model, Opus 4.8, with the user notified. A support page separately confirms that queries touching on elements of large scale model development may still get handled by that earlier version rather than the current one.

Anthropic is also working with Amazon, Google and Microsoft on a shared framework for scoring how dangerous a given AI jailbreak actually is. The draft weighs four factors:

• Capability gain, how far a technique pushes a user beyond tools they could already access.
• Breadth of capability gain, how many distinct offensive tasks it unlocks at once.
• Ease of weaponization, how much additional human effort turning it into an actual attack still requires.
• Discoverability, how easily someone else could stumble onto the same technique.

None of that framework was in place when Fable first locked down in June. It is the reason developers are still hitting refusals on routine coding and cybersecurity questions today.

The Token Capital Complaint

Nadella did not stop at Fable’s refusals. He used the same meeting to argue against an AI market run by a tiny handful of labs.

“It can’t be that there are only two companies in the world with token capital, and everybody else is renting it,” he told the engineers. “It makes no economic sense.”

Tokens measure how much computing an AI model burns through. Nadella’s framing turns Anthropic and OpenAI into landlords and everyone else, including Microsoft’s own enterprise customers, into tenants. It is a pointed thing for the CEO of a $3 trillion-plus cloud company to say out loud, given Microsoft’s own history of leaning on a single AI supplier.

Microsoft tied itself tightly to OpenAI for years before the two companies drifted apart following the abrupt 2023 ouster and reinstatement of OpenAI chief executive Sam Altman, a move Nadella reportedly learned about with little notice. OpenAI said in April it would extend its models beyond Azure to Amazon Web Services, ending Azure’s exclusivity. Nadella has since restructured Copilot itself around the lesson, putting former Snap executive Jacob Andreou in charge of unifying the consumer and enterprise versions, something he said this week “we should have done maybe day one.”

Microsoft’s Own Models Are Already Shipping

The independence Nadella is describing is not hypothetical. Microsoft’s AI division, led by Mustafa Suleyman, has spent the past six weeks pushing seven in-house MAI models into production, according to Microsoft AI’s own launch announcement. The flagship, MAI-Thinking-1, is Microsoft’s first reasoning model trained without any distillation from a third-party lab, including OpenAI’s.

Microsoft says MAI-Thinking-1 matches Anthropic’s own Opus 4.6 on coding benchmarks, at what it describes as a lower token cost. That is the direct rebuttal to Nadella’s “renting” line. If it holds up in enterprise use, Microsoft customers get a Microsoft-owned alternative sitting right next to Fable in the same Copilot model picker. Microsoft’s five-year push toward AI independence did not start with this week’s comments. It started at Build in June, and the Fable dispute is the first real test of whether that bet was timed right.

A Partner Microsoft Still Can’t Quit

None of this makes Anthropic disposable to Microsoft. The two companies are still deeply entangled financially. Microsoft committed $5 billion to Anthropic in November, while Anthropic pledged $30 billion toward Microsoft’s Azure cloud over the same period, part of a broader alliance that also saw Nvidia commit up to $10 billion of its own to the startup. Microsoft launched Copilot Cowork this year as a workplace assistant built directly on Anthropic’s technology, and Anthropic’s Claude Code tool has become one of the most widely used AI coding products among professional and amateur programmers alike.

Anthropic has separately said it needs more Microsoft compute, not less. Chief executive Dario Amodei said earlier this year the company has had “difficulties with compute” as Claude and Claude Code adoption outpaced its own infrastructure. That dependency runs in both directions, which is part of why Nadella’s public criticism of a company his own balance sheet is tied to reads as unusual even by AI industry standards, where partners and competitors routinely overlap.

Is Microsoft Actually Walking Away From Anthropic?

No, not yet, and probably not soon. Microsoft is building its own MAI models as an alternative inside the same Foundry catalog that still carries Anthropic’s Claude lineup, giving customers a choice rather than replacing Fable outright. The $30 billion Azure commitment and the Copilot Cowork product built on Anthropic’s models remain in place even as Nadella criticizes the safeguards publicly.

What is changing is leverage. Microsoft’s Foundry service now offers developers access to more than 11,000 models, including options from both Anthropic and OpenAI, alongside its own. That marketplace positioning lets Nadella criticize Fable’s refusals in public while still collecting Anthropic’s Azure spending in private, a balance that would have been harder to strike before Microsoft had a credible in-house model to point to.

The pressure on that balance is building from outside, too. On Thursday, Chinese startup Moonshot AI unveiled an open-source model it says outperforms recent releases from both Anthropic and OpenAI, adding a third competitive front just as Microsoft tries to diversify away from the other two. Anthropic, meanwhile, is reportedly preparing to meet with investors ahead of a possible public listing as soon as October, a milestone that would put fresh scrutiny on exactly the kind of restrictions Nadella spent Wednesday’s meeting complaining about.

Microsoft shares were down 2% in early Friday trading, a move investors tied more to broader worries about AI coding disruption than to Nadella’s comments specifically.