If you are building an app on Claude and you do not have a defense contract, your service is fine. Today. That is the most important sentence in this article, and it has a tail of caveats that every non-developer shipping on Anthropic should understand by the end of the week.
Here is what happened. On February 27, 2026, President Trump directed every federal agency to cease use of Anthropic's technology, with a six-month phase-out for the Department of Defense (CBS News). On March 3, the Pentagon — operating as the Department of War — sent letters formally designating Anthropic a supply chain risk. It was the first time a US company had ever received that label, normally reserved for foreign adversaries (NPR; Mayer Brown). On March 9, Anthropic filed two suits — one in the Northern District of California, one in the DC Circuit Court of Appeals — challenging the designation as retaliatory and procedurally defective (Lawfare). On March 26, a federal judge granted Anthropic a preliminary injunction, calling the designation "classic illegal First Amendment retaliation" (CNBC).
Anthropic is not going away. They are litigating. But "litigating" is not "stable." A reseller's policy could shift, a court could change scope, your customer's procurement team could panic. Run this audit today.
Open the file or the no-code block where your app calls Claude. Where does the request go?
api.anthropic.com — direct Anthropic APIbedrock-runtime.<region>.amazonaws.com)<region>-aiplatform.googleapis.com)Claude is the only frontier model available on all three major clouds, and AWS, Google, and Microsoft all confirmed publicly on March 6 that Claude remains available to non-Defense customers through their managed services (TechCrunch). If your app routes through Bedrock or Vertex, you are insulated from a lot of the legal turbulence — your contract is with Amazon or Google, not Anthropic. If you are calling Anthropic's API directly, that is fine, but it is one more thread in the dependency.
Action today: find the SDK call. Note the endpoint. Write it down.
This is a 2026 fundamentals question. If your app dies when one provider blinks, you have a single point of failure regardless of which provider it is. The minimum bar:
If you cannot point at the file that does this, that is the work for tomorrow.
Open your system prompt. Does it use Anthropic-specific tool-use schema (<tool_use> XML tags, the Files API, computer use, prompt caching breakpoints)? Those features do not transfer cleanly to OpenAI or Google. The more your prompt assumes Claude's idiosyncrasies, the more brittle a forced cutover becomes.
Action today: strip the prompt down to model-neutral instructions. Move tool definitions into JSON-schema function calls, which all three major providers support. Keep the Anthropic-specific tweaks in a thin adapter layer.
If you have to switch providers under time pressure, what does your retention exposure look like?
/messages endpoint with metadata.user_id setAnthropic's enterprise data policy does not retain customer API inputs for training by default, but you should know exactly what you have sent and where the artifacts live. Run an export of your last 30 days of API logs to your own storage now, so you are not asking for them later.
Three features have no clean cross-provider equivalent today:
Each of those is great until you need to leave. Audit your code for usage. If it is everywhere, that is technical debt with a regulatory tail. If it is in two clearly-marked functions, you are fine.
If you are on Anthropic's monthly API credit card billing, you are at the mercy of a Terms of Service that can change with email notice. If you are large enough to have an enterprise agreement, pull it up and search for "termination," "force majeure," "successor," and "service continuity." If those clauses are weak, your renewal is a good time to fix them. Reseller contracts (Bedrock, Vertex) carry the same clauses you negotiated with the cloud, which is usually stronger.
Run this checklist as a one-page document. For each item, write the answer, the file path, and the date you checked. Re-run it after every Anthropic news cycle — they are coming roughly weekly right now. The Mayer Brown and Northeastern coverage are useful trackers (Northeastern; Bloomberg).
You do not need to leave Claude. Claude is, by most evals, still the strongest coding model on the market. You need to know exactly what you would do if you had to. That is the difference between a side project and a real piece of software.
No. Per AWS, Google, and Microsoft's joint statements on March 6, Claude remains available to all non-Defense customers through their managed services. The Pentagon designation, the Trump directive, and the lawsuits all currently affect Defense Department and federal-agency use only.
Contractually, a lot. Direct Anthropic billing is governed by Anthropic's Terms of Service, which can change with email notice. Bedrock and Vertex contracts are with AWS and Google respectively, and carry the termination, force-majeure, successor, and service-continuity clauses you've already negotiated with the cloud — usually substantially stronger.
No. Claude is, by most evaluations, still the strongest coding model on the market, and a March 26 preliminary injunction in the Northern District of California is keeping civilian agency use in place. The point of the audit is to know exactly what you would do if you had to switch — not to switch.
Discover more content: