NVIDIA and Major Tech Companies Block Anthropic’s Fable Model Over Data Retention Policy

Anthropic's top investors are collectively pulling the plug.

NVIDIA, Palantir, and Booz Allen—some of the best-funded and most aggressive AI spenders in Silicon Valley—recently did the same thing in unison—

In core operations, completely block Anthropic’s flagship model, Fable.

Each party is more ruthless than the last.

A company that powers millions of American households was set to have Fable test the core grid, but Anthropic’s statement of “no guarantees” caused the plan to fall through immediately.

NVIDIA has replaced the entire valuable supply chain AI with its own Nemotron, leaving Fable to work only on insignificant open-source code.

Defense giant Northrop Grumman completely disconnected from the network physically and runs open-source models in its on-site data center.

The investors collectively turned against them, following Anthropic’s unilateral withdrawal of a promise made three months ago.

It all began on June 9, the day Fable 5 was released.

On that day, Anthropic quietly added a new rule to its support page—

All Mythos-tier models (i.e., the two strongest models, Fable and Mythos) retain user prompts and model outputs for 30 days for reasons of "trust and safety."

According to reports, conversations flagged as suspicious by the security system may be retained for up to two years.

Additionally, this policy applies to all existing Zero Data Retention (ZDR) agreements, with no toggle—it is enforced mandatorily.

Zero Data Retention is the most critical "life-saving clause" that enterprise customers prioritize when negotiating with AI vendors.

This means that anything I send you must be deleted immediately after processing—do not store it, do not view it, and certainly do not use it to train future models. Financial, medical, and defense companies would never accept an API without this requirement.

Before June 9, Anthropic signed ZDRs with several major clients, allowing Opus 4.8 and Sonnet 4.6 to use this expedited channel.

But after that, only the flagship Fable model won’t work. If you want to use the strongest model, turn on the data retention switch first—no exceptions.

Anthropic explained that the Mythos-level models are too powerful, and to defend against sophisticated new attacks—such as advanced jailbreaks that span multiple requests—the system must retain logs for monitoring.

This argument may sound reasonable, but Silicon Valley exploded in outrage that same day.

Jun Park, CEO of startup Hillclimb, summarized on X: “Anthropic’s new policy is that they collect your data whenever you use Fable or Mythos—no exceptions, not even for enterprise partners.”

Microsoft acted the fastest.

Less than a day after Fable 5 was released, Microsoft restricted internal employee access and required its legal department to review the retention terms first.

As a result, Fable is the only model missing from Microsoft's internal Copilot model list.

The first to publicly challenge was Palantir.

As a platform distributing AI models to governments and large enterprises, Palantir executives spent considerable time negotiating with Anthropic to obtain an irrevocable ZDR commitment covering all models, to reassure customers.

The new policy in June effectively nullified this commitment.

Palantir CEO Alex Karp directly criticized them. In early July, he spent a full 20 minutes scolding the large model labs.

At last week's customer conference, he openly stated that companies have had enough of being "abused" by AI labs.

Palantir even gave each attending client a booklet titled "How to Avoid Giving Away Your Alpha (Core Advantage) to托管Model Providers," guiding enterprises step by step on how to negotiate against ZDR clauses with AI vendors.

According to insiders, Palantir’s position is clear: Fable won’t be allowed on Palantir’s platform until Anthropic provides an irrevocable ZDR.

The customer wants to? Go buy it directly from Anthropic.

Turned around, Palantir served up OpenAI’s GPT-6 Astra on the table—because OpenAI gave ZDR.

Yes, just as Anthropic was still figuring out how to calm down these angry major clients, competitors struck with a knife.

On August 19, OpenAI publicly announced it is testing a system called "Private Safety Processing."

The logic of this system is the exact opposite of Anthropic's—

Customer data remains on the customer’s own infrastructure or is encrypted using the customer’s own keys; OpenAI receives only a minimal set of narrowly defined "security signals" and never accesses original prompts or responses.

Abuse will still be detected, but no data will be retained; the latest models continue to support ZDR.

This strike was extremely precise.

Anthropic just said, "Without retaining data, you can't ensure security," but OpenAI immediately proved otherwise—security is still possible without retaining data.

Just over ten days later, on September 1, Anthropic finally couldn't sit still anymore.

With the release of Fable 5.1, Anthropic introduced a set of Enterprise Frontier Safeguards (EFS).

The logic of this solution is almost identical to OpenAI's—

Data is stored on the customer’s cloud, with the customer retaining full control of the keys; Anthropic’s system is solely responsible for scanning for abuse signals and forwarding them to the customer for review. The company states that this solution was co-developed with over 100 customers across financial, healthcare, and telecommunications sectors, including multiple CISOs.

Anthropic even subtly admitted fault in its blog, acknowledging that companies in regulated industries "indeed find it difficult to use models that require data retention."

After a three-month standoff, it appears on the surface that Anthropic has conceded.

But when major clients examined the contract closely, they realized things weren't so simple.

The EFS program is currently invite-only, will roll out in phases this fall, and is not compatible with Mythos 5.1. Worse still, the terms include a backdoor: Anthropic can cancel this option at any time.

This statement directly crosses the line that Palantir fiercely insists on—the word "irrevocable."

In the eyes of legal professionals, a promise that can be revoked at any time is no promise at all. Today, a company signs an EFS and feeds its core code into it; tomorrow, Anthropic issues a notice canceling it, and the data processing procedure reverts to the old rule of retaining data for 30 days.

Enterprises want to hold the key in their own hands; what Anthropic provides is a key that can be revoked at any time.

As a result, on September 14, an extremely poor-looking "restricted list" was released.

Booz Allen CTO Bill Vass put into words the concern of every large enterprise: "We're a little worried that Fable will copy our code."

Yes, a paper contract can prevent “use for training,” but it absolutely cannot stop the model from thoroughly absorbing your most critical business logic and most valuable trade secrets within the 30-day retention period.

The most ironic thing is that both sides in this covert struggle are waving the banner of "security."

Anthropic insists on retaining data to prevent advanced models from being misused by malicious actors; large corporations refuse to retain data to avoid having their core code secretly copied.

Safety meets safety, leading straight to absurdity—the stronger the model, the more companies fear feeding it real data, as if guarding against thieves.

The world's most powerful flagship model can only handle a bunch of trivial, meaningless tasks.

According to Ramp’s survey of over 70,000 U.S. companies, two months after Fable 5’s launch, its spending share remained firmly at 11%, while the token share of open-source models surged from 11% to 62%.

But the idea that "you can control yourself" may also be a deadly illusion.

At the height of the argument, Anthropic released a threat intelligence report, prominently showcasing how it had been catching hackers in the background over the past several months.

A group of hackers repeatedly modified exploit code using Claude until they succeeded through trial and error—all of which Anthropic observed clearly.

Another group attempting to swap the API key with the victim’s was caught in the act.

Who, when, and what actions the model performed—clearly listed one by one.

Anthropic's intention was to flex its muscles; after reading it, the big corporations felt a chill down their spines—

If they can observe every step the hacker takes so closely, aren't we essentially "naked" in front of them?

In fact, the three-month debate over “whether to keep the data” may have been merely superficial. The real,致命 issue is the extent of your model provider’s godlike perspective.

Anthropic provided all major clients with a chilling answer through this hand-released report.

We can see everything you have.

This article is from the WeChat public account "New Intelligence Yuan," authored by ASI Revelation.