HHS sole-source award to Anthropic continues to raise questions

The Department of Health and Human Services shared new details about its plans to provide an award to Anthropic without a competition amid questions about its rationale.

In response to a FedScoop inquiry, an HHS spokesperson said the award will be “a temporary six-month sole-source bridge contract” and was necessary to “to maintain continuity of vital” AI services when a discounted offering from the Claude maker expires at the end of the month.

Like other major AI developers, Anthropic offered access to its tools through the General Services Administration’s OneGov program for $1 for one year. That deal, however, is expiring Oct. 31. OpenAI announced a deal for consumption-based pricing through 2028 before its expiration, but Anthropic has not yet made an equivalent announcement.

“During this interim period, HHS is actively pursuing a long-term contract to support the HHS [super intelligence] needs,” the spokesperson wrote in a Thursday email, using the administration’s self-styled term for AI. “Our objective is to have these longer-term arrangements fully executed and in place prior to the conclusion of the six-month bridge window.”

FedScoop also asked whether a solicitation for that long-term contract was public and for additional details about justification for the use of a sole-source award, but the spokesperson did not respond to those questions.

The response comes after the agency announced plans for the sole-source award earlier this week, and adds new information about the agreement’s length and role as a stopgap following the end of the OneGov deal. For those watching AI contracting, however, it still doesn’t address previous concerns that the approach could be a misuse of the sole-source process.

Jessica Tillipman, associate dean for Government Procurement Law Studies at George Washington University, expressed concern after the initial notice and said the additional details make things worse. 

“I think this does not help their case and, in fact, weakens it,” she said of the additional details.

The agency’s rationale for the sole-source award appears to be a circular argument that the only company that can provide Claude services is Anthropic, Tillipman said. They still haven’t explained why Claude is the only provider that could meet the requirements they outline — which describe general chatbot functions.

Knowing the sole-source award is a bridge contract after the OneGov deal raises new questions about why the agency was unable to prepare for the transition when the end date for that agreement was public, Tillipman said. The deal had also already been extended by a month.

Protest possible

David Timm, a partner at Burr and Forman focused on government contract litigation, similarly pointed to potential gaps in the agency’s justification. In addition to the questionable uniqueness of the services, HHS previously halted use of Anthropic amid a governmentwide ban, potentially weakening its argument about continuity needs, he said.

Following a falling out between the Pentagon and the Claude maker earlier this year, President Donald Trump issued a governmentwide ban on use of Anthropic’s tools. That ban has since been enjoined by a federal judge, but it initially prompted agencies, including HHS, to disable their products at the agency.

Amid that loss of continuity, HHS was “able to muddle through — presumably by relying on other models that they had access to. So I think that argument is also difficult to make,” Timm said.

A challenge might find footing on those points.

“Uniqueness and continuity — those are kind of tough sells here based on what we know from the notice,” Timm said. If another provider were interested in protesting the award, he said “there would be some there there.” Though, he added, the odds of success are difficult to determine without an actual record. 

Another consideration is whether there’s a vendor that also provides the services described that wants to challenge this case, Timm said. There are “a variety” of vendors that might be interested parties, such as OpenAI or Google. But if they’re already selling to HHS, “do they care if this sole source goes to Anthropic here because they’re already providing their services?” he said.

Award amount

Interestingly, the legal authority HHS used for the award may also provide some insight into the contract amount. 

The award was made under a part of the Federal Acquisition Regulation (FAR) known as the simplified acquisition threshold, which permits a more streamlined process for contracts under a specified dollar amount. If appropriately made under that authority, the contract would be less than $350,000, which is the current threshold limit. 

Timm noted it’s not clear whether that was the appropriate authority because the agency does not provide cost in its notice. If contracts are above that limit, however, they require more paperwork, including a formal justification and approval document, which the notice does not have.

Lock-in concerns

Concern about so-called “lock-in” — or agency dependency on particular vendors — has been a part of the AI conversation. Tillipman notably discussed the issue in a January paper on AI procurement risks, and it’s recently come up in conversation among government officials as well.

At a FedInsider event earlier this week, Natalie Buda Smith, chief AI officer and director of digital strategy at the Library of Congress, talked about a need to ensure that organizations remain flexible with their AI use.

“Some of these tools are great, and they offer multiple tools within,” Buda Smith said. But something to consider with procurement is “not locking in to one way of doing things.” She added that organizations need flexibility because models change and “you might have an approach and you need to extend it.”

Buda Smith used the analogy of budgeting for a property to illuminate the issue. Where pilots are a down payment, the budget also includes a mortgage, utilities and property taxes when the project eventually moves to production.

“We tend to buy AI tools as capital, but they are really utilities,” Buda Smith said. “So it’s a whole different way of thinking about this.”

K. Sophie Will contributed to this article.