Will Anthropic and OpenAI Stop Selling Their Best AI to Businesses?

Anthropic and OpenAI’s move to develop AI applications and features tailored to specific industries continues to worry their customers, some of which must now compete head-on with their main AI suppliers. Design app Canva is the latest to find itself in this predicament, as we reported earlier this month.

Through it all, though, most people assumed Anthropic and OpenAI would always offer their models through an application programming interface, which is a multibillion-dollar revenue driver for both of them.

As long as it stays that way, businesses will stand a chance against the new competition. But will it?

Conversations I’ve had in recent weeks with researchers at these frontier labs and founders whose applications depend on the latest models left me wondering if the status quo will persist.

There are more than a few signs that Anthropic and OpenAI could degrade API access to their best models, whether it’s their choice or the government’s. The Trump Administration has pushed these companies to pursue a more staggered model-release process for some recent models due to concerns that they might be used by cyber criminals or other bad actors. During this period, customers have had to be approved on a case by case basis. (The Trump administration recently previewed a special testing process major AI firms could follow before releasing models.)

And even after releasing models publicly through an API, Anthropic and OpenAI may degrade the quality of their models’ performance on some tasks for similar safety reasons. They have already done so in some cases, like in the case of Anthropic’s Fable with cybersecurity-related tasks.

Of course, Anthropic and OpenAI wouldn’t necessarily have to degrade performance when they use their models to power their own specialized apps or features.

Both companies have warned that future models could become dangerously capable in areas like hacking and developing bioweapons. At some point, the AI providers could decide it is too difficult to try to block every attempt by bad actors to use models for these purposes; if criminals do break through, the blowback could be severe. As crazy as it sounds, they may decide it’s safer to cut off access to advanced models altogether.

Anthropic and OpenAI have other reasons to restrict API access to such models. Both of them want to prevent distillation, a technique that allows competing AI firms, particularly from China, to train new models in part by using answers or outputs from advanced models that are commercially available.

One former Anthropic researcher told me it’s basically impossible to completely stop distillation. And if Anthropic and other American AI developers are truly worried about bad actors using models to wreak havoc on the world, they probably won’t feel great about indirectly helping Chinese firms to develop advanced, open-weight models that are available for free to anyone.

As Anthropic noted last week in its report on the potential catastrophic risks of its AI, “if the resulting distilled model inherits risk-relevant capabilities of the original model but is served without adequate safeguards, this could impose risk on the world downstream of the original model’s capabilities even if the original model had strong safeguards against direct misuse.”

For their part, open source software advocates say freely available models, including those distilled from Anthropic models, can help developers protect themselves against such cybersecurity attacks in particular, as was the case for Hugging Face, which was hacked by OpenAI agents.

Then there’s the most cynical reason Anthropic and OpenAI might have to cut off API access. As we covered in this piece about Anthropic blindsiding its business partners, some investors have been warning developers that Anthropic could hold back its best technology so it can develop its own competing applications.

Suffice it to say that there would be a mega antitrust investigation if Anthropic, as the AI API market leader (by a wide margin), decides to do this.

Of course, OpenAI and Anthropic probably wouldn’t be thrilled to give up a major source of revenue with their API. However, AI-powered applications could provide a potentially even more lucrative source of revenue. Anthropic is even getting into fields like AI-powered drug discovery.

For now, the uncertainty around future API access explains why AI app developers such as Harvey and Cursor have taken steps to train their own in-house models rather than depend on OpenAI and Anthropic. Others are sure to follow in their footsteps.

Here’s what else is going on…

Overheard

Anthropic CEO Dario Amodei pushed back on the idea that he has been overly pessimistic when it comes to AI in a X post on Saturday.

“I do not agree that my messaging has been disproportionately negative,” Amodei wrote. “In fact it has been about equally balanced between risks and benefits: I’ve written one major essay about each, and even in interviews where I discuss the risks, I make sure to frequently mention the incredible benefits as well as proposing possible solutions to the risks (short clips from my interviews that end up on social media tend to be disproportionately negative, as that gets clicks).”

People on the Move

OpenAI said Thursday it is hiring Dali Rajic, president and chief operating officer of Alphabet-owned cybersecurity firm Wiz, to replace Denise Dresser as its new chief revenue officer. Dresser is departing the company “to pursue other opportunities following a transition period,” according to a blog post. Dresser, the former CEO of Slack at Salesforce, is leaving after just eight months.

Tarun Chhabra, a former Biden administration official who serves as Anthropic’s head of national security policy, is transitioning out of his role, The Information reported on Thursday. Chhabra will take on a new advisory role at Anthropic focused on geopolitics and national security policy and will continue to work with Anthropic leadership. It’s unclear if he will remain full-time.

The move comes amid a broader restructuring of Anthropic’s government affairs operations, which will now sit under Mariano-Florentino Cuéllar, who Anthropic announced as its new chief global affairs officer earlier this month.

Deals and Debuts

See The Information’s Generative AI Database for an exclusive list of private companies and their investors.

SpaceX completed its $60 billion acquisition of coding startup Cursor on Friday, the companies said.

Anthropic’s revenue rose around 14 times in the second quarter from the same period a year ago. That brought total revenue to $11.5 billion, from $787 million in the year-ago quarter and $4.73 billion in the first quarter, according to Bloomberg, which first reported the figures. It was also profitable on an adjusted operating income basis, the report said.

Databricks, a data analytics and AI company, raised $5 billion in strategic funding at a $190 billion valuation led by Coatue Management. The company announced that it has passed $7 billion in annualized revenue.

Thrive Capital’s 2022 fund—a $516 million vehicle that invested in SpaceX, OpenAI, Cursor, Anduril, Ramp, Wiz and Databricks—has risen in value more than seven times net of fees as of June 30. The multiple of total value to paid-in capital, which includes paper and cash returns, puts Thrive’s fund well above top decile funds of the same vintage year, which have a TVPI of 1.73 times on average, according to PitchBook benchmarks as of year-end 2025.

Silicon Data, a New York City-based startup that offers data and benchmarks on compute, raised $30.5 million in Series A funding led by Valor Atreides AI Fund.

Mindgard, a Boston-based AI security software company, raised $30 million in Series A funding led by Album VC.

Chinese AI developer Z.ai on Friday released its new open-source model, GLM-5.3, saying its cybersecurity capabilities are on par with Anthropic’s Mythos 5.

DeepSeek announced a significant increase in the prices of its models, a day after it released its flagship model, V4-Pro.

Google announced on Friday that it will now allow users to remove a visible watermark from its AI-generated content, including images, videos, and songs.

OpenAI announced a new mode called Ultrafast, which aims to run GPT-5.6 Sol up to 14 times faster than normal. The new mode is powered by Cerebras chips, the company said.

Thank you for reading the AI Agenda Newsletter! I’d love your feedback, ideas and tips: [email protected].

If you think someone else might enjoy this newsletter, please pass it forward or they can sign up here.