Anthropic And OpenAI Back 'Pacing' Frontier AI. What Happens To Enterprise AI Plans?

Anthropic CEO Dario Amodei called for more deliberate pacing of frontier model development, an idea endorsed by OpenAI CEO Sam Altman and xAI founder Elon Musk. But what does "pacing" mean for organizations already building AI into their operations?

The leaders building some of the world's most powerful artificial intelligence systems are publicly entertaining a proposal that seems almost extraordinary as enterprises race to adopt the technology: It may be time to deliberately slow the pace of frontier AI development.

In an essay published Saturday, Anthropic CEO Dario Amodei argued that AI companies need to give safety research, testing and external oversight time to catch up with the rapidly advancing capabilities of models. Subsequently, OpenAI CEO Sam Altman endorsed Amodei's suggestion and said his company would match Anthropic's proposed commitment to giving independent evaluators employee-like access. Elon Musk, founder of xAI, also backed Amodei's position.

In his essay, "We Must Pace The Frontier," Amodei does not call for halting model development. Instead, he defines pacing as slowing capability development enough "so that risk prevention has time to keep up."

His three-part proposal begins with placing independent evaluators inside frontier companies and giving them access as though they were employees conducting risk assessments. He then calls for AI companies in democratic countries to coordinate on common safety standards and limits on unchecked progress, followed by international coordination efforts.

What Amodei does not address is what pacing could mean for enterprises that have invested heavily in incorporating AI tools into their operations.

From Frontier Safety To Enterprise Planning

Dario Amodei identified two major concerns in his essay. The first is AI's growing role in developing future AI systems, a process known as "recursive self-improvement." The second is the OpenAI-Hugging Face incident, in which AI agents reportedly conducted cybersecurity attacks beyond their assigned task and attempted to compromise the system responsible for evaluating their performance.

Amodei argues that the incident demonstrated how a more capable group of agents could potentially wreak far greater damage. He contends that AI's dizzying rate of advancement warrants giving security, alignment, interpretability and testing more time to keep up with model development.

(Anthropic CEO Dario Amodei. Image credit: Anthropic)

IT leaders share many of the same concerns about AI security, control and governance. But any significant changes to the development cycle of the AI models that now power many critical business applications would raise several important questions for enterprise customers.

Will there be longer periods between major frontier-model releases? Could slower development by leading U.S. AI companies give malicious actors or overseas competitors an advantage? Will organizations require more extensive model evaluations before deployments? Could vendors change product road maps, preview programs or expected release schedules?

Perhaps the most immediate question is whether existing enterprise products would continue to operate normally even if the development of the most used frontier models in businesses were paced.

Does Oversight Equal Pacing?

David Lindner, CISO, Contrast Security, said Anthropic's first commitment does not, by itself, amount to an actual reduction in development speed.

"Anthropic is asking the whole industry to slow down, but it's not actually slowing down itself. It's just adding somebody to watch it not slow down," Lindner said in an emailed statement to MES Computing.

He compared the situation to the history of vulnerability-disclosure practices, when vendors were reluctant to act first without rules applying across the market.

'We've seen this before in security," he said. "Vendors sat on vulnerability disclosures for years because nobody wanted to move first, and real regulation was required before anything changed."

If the risks described by AI leaders are as serious as they suggest, Lindner said companies should be willing to accept hard limits rather than wait for competitors or lawmakers to officially set them.

(David Lindner, CISO, Contrast Security)

"If the risks are as serious as Dario and OpenAI and others say, "I'd rather see Anthropic commit to an actual pace limit now instead of waiting on Congress or a competitor to go first," he said.

Lindner's scrutiny of the proposal gets to the heart of the debate. While Anthropic says it is committed to more oversight, it has not committed to actually slowing the pace of AI development. Third-party evaluators may provide more transparency into how frontier AI systems are developed and tested, but they cannot set limits on how quickly new capabilities are created or released.

While Amodei suggests transparency as the initial step in his proposal, it remains an open question whether oversight without real set speed limits translates into pacing AI at all.

Pacing Could Deliver More AI Stability To Organizations

As successful business outcomes grow more dependent on a technology that can change substantially from budgeting to deployment, slower frontier AI development may prove beneficial, Lindner argued.

"Right now every function betting on AI is stuck planning around a moving target," he said. "A new model ships, capabilities jump, and legal has to re-review what the tool is allowed to touch, procurement has to re-run vendor risk, finance has to redo the ROI case because the thing they budgeted for already changed."

Lindner said that a more deliberate development cadence could help organizations better understand the capabilities of AI models, and thus, better governance over them.

"Pacing shrinks that gap between what a model can suddenly do and what the rest of the business has actually had time to adapt to," he explained.

Security benefits may be most apparent. Slower capability growth could not only lead to better governance but may also give monitoring programs a chance to keep pace. Third-party evaluators may offer IT leaders more substantive evidence to conduct reviews rather than rely solely on AI vendor assurances.

Lindner suggested that the potential benefits could extend throughout an organization.

"Procurement gets more stable contracts to write, legal gets more time to actually understand what they're approving, and finance gets a forecast that doesn't get blown up every few weeks by a capability jump nobody added to the budget," he said.

Those scenarios could be particularly valuable for midmarket organizations. With smaller legal, security, procurement and governance teams than their large-enterprise counterparts, they may have less capacity to repeatedly assess AI products as underlying models, capabilities and risk changes.

Existing AI Would Not Simply Disappear

Pacing does not mean pausing AI development. Amodei said that a pause "is unlikely to actually happen any time soon" and that "defecting from such an agreement by evading monitoring could radically shift the balance of global power," and no frontier AI provider has announced plans to withdraw currently available models as part of this debate.

"People hear pause and picture their AI going away, but no one is suggesting pulling access to models that are already available," Lindner said. "If you built your business on today's models, you keep everything you built. Pacing doesn't touch that."

Neither Anthropic nor OpenAI provided MES Computing with details about how pacing might affect existing products, support contracts, or model availability.

A greater risk may be for organizations whose business cases depend on anticipated improvements to current models, Lindner explained.

"Where [pacing] actually may create a slight problem is the organizations that banked on next year's models being dramatically better, agents that finally work reliably, reasoning that finally holds up, and built the business case around a capability jump that hasn't happened yet," he said.

"Slower frontier progress is fine if you already got value out of what you have. It's a real problem if your plan was always waiting on the next model to make the math work."

Microsoft Points To A Human-Controlled Model Framework

MES Computing contacted Anthropic and OpenAI about Amodei's proposal and its implications for enterprise customers but did not receive responses by publication time.

A spokesperson supporting Microsoft’s MAI and CoreAI teams declined to comment directly but referred MES Computing to a blog post published Monday outlining the company's developing approach to AI safety.

In "Humanist AI in practice: A public consultation on our Code of Conduct for MAI Models"Microsoft AI introduced a draft code intended to define how its MAI models should behave, the values they should follow and the boundaries they must not cross.

According to the blog post, Microsoft's code is "designed to ensure MAI models will never resist human interruption, correction, or shutdown. That they will not widen their own scope, take on goals no human has given them, or hide their reasoning from the people auditing them."

The code of conduct is not currently being used to train MAO models, according to Microsoft AI. The company describes the document as a "work in progress" that will eventually serve as a primary governing document for model training, technical controls, operational monitoring and evaluation.

Microsoft's code addresses some of the concerns Amodei raises in his blog post, including the need for human oversight, clear behavioral limits, and the ability to interrupt or shut down AI systems.

However, as with Amodei's post, MAI's code does not answer the questions this debate creates for its customers: Can a slower frontier-model cadence affect Microsoft Copilot product road maps, model availability, or capability improvements delivered via updates?

Enterprise AI Plans Need Reality Checks, Not Retreats

At this stage, there is no agreed upon, industry-wide slowdown, no common definition of what an acceptable development pace even means, and no timeframe for putting a plan such as Amodei suggests into effect.

Organizations therefore have little reason to halt AI projects while frontier AI model developers ponder the pace of development.

IT leaders should, however, determine whether an AI investment provides business value from AI model capabilities available right now, or whether the business case depends on future models solving technical and operational problems that existing AI tools cannot.

They should also ask vendors and providers how changes in development cadence could affect product road maps, which underlying models are powering their services, and whether customers are notified when safety findings may impact functionality or availability.

For midmarket organizations, the most immediate concern is not that pacing will make exiting AI disappear. It is that organizations may have built long-term business plans around AI capability improvements that AI industry leaders are now saying may be developing too quickly.