comma.ai bans Anthropic as an AI vendor after CTO warns of cloud token dependence

Why it matters

Schaefer is treating hosted AI coding tools as supply-chain risk, forcing the procurement question many startups skipped while racing to adopt Claude and similar tools.

comma.ai CTO Harald Schaefer (@Harald) said on July 7th that Anthropic is now a banned vendor at the autonomous-driving software company, turning a vendor choice into a public warning about how much engineering teams have ceded to cloud model providers.

https://x.com/___Harald___/status/2074561342539956403

"Anthropic is now a banned vendor at @comma_ai," Schaefer wrote on X. "I recommend other companies do the same." In the same post, he told companies to examine whether their employees can still brainstorm, code or stay productive without hosted token vendors.

The attached screenshot gave the internal rule more teeth. In a message addressed to comma employees, Schaefer wrote that they "are no longer allowed to send comma dollars to anthropic," citing two objections: Anthropic's treatment of user sovereignty and what he described as an effort to push business customers into high usage costs. The screenshot specifically references "API billing only for enterprise customers" and usage-credit mechanics around Fable, while arguing that comma.ai should spend time and money on its own infrastructure instead.

Schaefer's move lands differently from a standard procurement change because comma.ai already sells hardware built around local compute. The company markets the comma four as "an AI upgrade for your car" and says its hardware can run openpilot, the open-source driver-assistance system that comma.ai says works across more than 300 supported cars. The company also claims more than 300 million miles driven, 20,000 users and a GitHub repo with more than 50,000 stars on its homepage.

That makes Anthropic more than a software vendor in this fight. For comma.ai, whose public pitch has long centered on shipping usable autonomy from a smaller, vertically integrated team, dependence on a hosted coding assistant cuts against the operating model. Schaefer's note says as much: "Comma sells ML computers, we are capable of doing model inference ourselves as long as there are decent open models."

Schaefer has standing inside comma.ai to make that argument. UCLA described him in a 2024 talk listing as comma.ai's CTO and said he had worked at the company for more than seven years developing an end-to-end self-driving system for openpilot. The same listing described comma.ai's training approach as using a large world model that can generate driving video, allowing policy models to train in simulation before deployment in openpilot.

Anthropic's own documentation supports part of the billing backdrop Schaefer is criticizing. Anthropic says Claude Code uses Anthropic's API by default, and its help center says API access, Workbench usage and Claude Code can draw from usage credits. For Enterprise plans, Anthropic says usage-based plans charge usage separately from seat fees based on token consumption at standard API rates.

The unresolved part is Schaefer's broader claim that Anthropic does not respect user sovereignty. The X post does not specify the incidents behind that assessment, and Schaefer did not tie the ban to a single event in the public post. The internal note instead treats the vendor ban as an architectural decision: comma.ai should avoid making its engineering workflow dependent on a provider that can change pricing, product access or terms from outside the company.

That is the practical edge of the announcement. AI coding tools have become infrastructure inside engineering teams before many companies have decided whether those tools should be treated like productivity software, cloud infrastructure or a supply-chain dependency. Schaefer is putting Anthropic in the third bucket.

The choice also fits comma.ai's founder-era culture. George Hotz built comma.ai around a public, hacker-inflected rejection of the closed autonomous-vehicle model. Today, the company describes its GitHub organization as "Building the Android of self-driving cars" and lists public datasets including commaVQ, commaSteeringControl, comma10k and comma2k19. Schaefer's Anthropic ban extends that posture from cars to internal software production: if the team cannot work without a remote model provider, the dependency becomes part of the product risk.

There is a cost to that posture. Claude and similar hosted coding tools have become popular because they remove operational work from teams that would rather rent frontier inference than maintain it. Schaefer's answer is to spend the operational work anyway, using open models where they are good enough and comma.ai's own infrastructure where control matters.

For Anthropic, the immediate business impact of losing comma.ai as a customer is likely small. The reputational impact is more pointed. A deep-learning-heavy robotics company banning a frontier AI vendor over sovereignty and billing gives other technical buyers a sharper vocabulary for concerns that often sit below the surface in procurement reviews. The question procurement teams will ask is concrete: if an AI tool becomes core to how engineers write and reason about software, can the company replace it fast enough when the vendor changes the rules?