A court filing dropped this week in the Anthropic-Pentagon dispute, and the detail that jumped out was almost comically specific: internal communications show the two sides told each other they were “nearly aligned” on contract terms. One week before the entire relationship was killed. Seven days from handshake to blacklist, with both parties apparently surprised by how fast the ground moved.

I keep thinking about that one-week gap. Not because of what it says about defense contracting or AI policy, but because of what it reveals about how fragile every AI provider relationship actually is, including yours.

Seven days is not a warning period

Most people who build workflows around a single AI model operate on the assumption that they would get some notice if things went sideways. A deprecation timeline, maybe. An email saying “we’re changing our terms effective next quarter.” Something that gives you enough runway to find alternatives and migrate at a reasonable pace.

Anthropic’s own leadership apparently believed the Pentagon deal was on track. They were wrong by exactly one week. And if a company with $200 million on the line and direct communication with its counterparty could not see this coming, the idea that you or I would get meaningful advance warning before our provider access changes feels almost naive.

The provider relationship you trust is a snapshot

Every AI company is simultaneously negotiating with governments, investors, competitors, and regulators across dozens of jurisdictions, and any one of those conversations can cascade into product changes that affect you without anyone at the company intending that outcome. Anthropic did not wake up one morning and decide to lose federal access. OpenAI did not plan for half the internet to sue them. Google’s antitrust situation is not something the Gemini team chose.

But the downstream effects land on you regardless of intent. Your custom prompts, your carefully tuned workflows, the way you have learned to phrase requests so your preferred model gives you exactly the output format you need. All of it sits on top of a provider relationship that looks stable right up until it is not.

BYOK is a business continuity strategy now

I wrote about this when the Pentagon story first broke in March. At the time, the argument for Bring Your Own Key was mostly about privacy and data sovereignty. That argument still holds. But the court filing adds a dimension that I think matters more for most people: BYOK is how you avoid caring about provider drama entirely.

When the AI model is a swappable component underneath your tool rather than the tool itself, the question of whether Anthropic and the Pentagon can sort out their differences becomes genuinely irrelevant to your Tuesday afternoon. Claude acting weird after an update? Switch to GPT for the rest of the week. OpenAI raising API prices? Route your tasks through Gemini. None of this requires rebuilding anything because the interface layer — the part you actually interact with — stays identical.

Your browser already solved half this problem

And this is where the local browser piece connects, because BYOK only works as a real continuity strategy if the tool running the model can actually do useful work regardless of which model is active. A cloud-hosted AI assistant that deeply integrates with Claude’s specific API features will still break when you swap providers, even if it technically supports multiple models. The integration points become the lock-in.

Dassi runs as a Chrome extension that reads your actual browser page through DOM access. It does not depend on any provider-specific integration for its core capability, which is seeing what you see and acting on it. The model behind it handles reasoning and language. Your logged-in browser sessions, your cookies, your authenticated state — that context comes from Chrome itself, not from any AI company’s cloud.

So when you swap from Claude to GPT to Gemini inside dassi, the agent keeps reading the same page, keeps accessing the same logged-in apps, keeps operating in your browser with your data. The model is the one replaceable part in a stack where everything else — your sessions, your screen, your workflow — belongs to you.

Nobody’s provider relationship is special

I realize some people reading this have strong preferences about which AI model they use, and that’s fine. Claude is better at certain tasks. GPT handles others more reliably. Gemini has its own strengths. Having a preferred model and being locked into one are very different things, and the court filing this week is a reminder that the distance between “preferred partner” and “former partner” can be measured in days.

The Anthropic-Pentagon situation is dramatic because of the scale and the players involved. But the underlying dynamic — two parties who thought they had a stable relationship discovering they did not — plays out constantly in tech. APIs get deprecated. Pricing changes. Terms of service shift. And the people who built contingency into their setup barely notice, while everyone else scrambles.

A browser agent that runs locally with swappable models is not a hedge against some unlikely scenario. It is the baseline architecture for anyone who does not want to rebuild their workflow every time a provider has a bad quarter.