Two AI models stopped answering this month, for two completely different reasons, and only one of them came back. Google's Gemini went dark for a few hours during a technical outage, then recovered. Anthropic's brand-new Claude Fable 5 went dark three days after launch, pulled by the US government over a discovered jailbreak, and as of this writing it has been offline for eight days with no return date. Both events triggered the same quiet response from companies that had already prepared for exactly this: software switched people to a different AI model automatically, without anyone touching a setting.

The easy button built for exactly this moment

The Gemini workaround is a product called Cloudflare's AI Gateway. Think of it as a power strip for AI. Your apps plug into it instead of plugging directly into Google, OpenAI, or Anthropic. If one provider goes down, the gateway reroutes traffic to another one automatically. It also watches every AI request that passes through for security problems, and it lets a company set a spending limit so nobody on the team accidentally runs the most expensive AI model available on a task a cheap one could have handled just fine. Cloudflare's CIO, Sam Rhea, used the Gemini outage as the live example: teams using the gateway kept working through it without anyone touching their configuration.

Then a bigger one went dark, and it hasn't come back

Three days after Anthropic launched Claude Fable 5 as its most capable public model yet, the US Commerce Department ordered it pulled. The directive, issued June 12, cited national security concerns tied to a method for bypassing the model's safety controls. Because the order applied to any user outside the United States, and Anthropic had no way to check in real time who was logging in from where, the company disabled the model for everyone, everywhere, all at once. Every other Claude model kept working normally throughout. Anthropic has said it believes the situation stems from a misunderstanding and is contesting the order, but as of this writing, eight days later, Fable 5 is still offline and the company has given no timeline for when that changes.

Anthropic had already built a smaller version of this same idea into Fable 5 before any of this happened. If you asked it something that touched a restricted area, cybersecurity or biology, the model wouldn't answer directly. It would quietly hand the question to a different, more limited model instead, in fewer than five out of every hundred conversations. That feature is still part of how the model is designed to work. It just doesn't matter much right now, since the whole model is the thing that got switched off.

Routing isn't a future feature anymore. It's already running inside AI models you currently can't use.

A tech CEO says this switching is the whole business now

Most of what a company actually uses AI for, drafting emails, sorting documents, filling out forms, doesn't need the smartest model available. Box CEO Aaron Levie has been making this point for months: the businesses that win won't be the ones with the single best AI model. They'll be the ones that can tell, automatically and on every request, which job needs the expensive model and which job doesn't. His argument gets harder to dismiss now that the "best" model can vanish within days of becoming available. If your strategy depends on routing the hardest problems to the single smartest model on the market, what happens the week that model is the one a government just turned off?

The most expensive AI on the market is now built to avoid using itself when it doesn't have to, and built to disappear entirely when a government tells it to.

Some companies already sell this as a ten-dollar subscription

Routing doesn't only show up as enterprise infrastructure. Abacus.AI sells a product called RouteLLM that does the same basic job Cloudflare's gateway does, automatically pick a model based on the request, fail over if a provider goes down, for ten dollars a month. I've used it myself as one of five AI subscriptions I run side by side, and the honest finding from testing it against a question with a known answer was that automatic routing solves cost and uptime, but it doesn't solve a model getting a simple, verifiable fact wrong. Routing picks you a model. It doesn't make that model right.

Here's the catch: even the test graders disagree

All of this assumes somebody can reliably tell which model is "good enough" for a given job. This week showed that's harder than it sounds. A testing group called Design Arena pitted a free, open model called GLM-5.2, built by a company called Z.ai, against Claude Fable 5 on building websites. GLM-5.2 won outright, beating Fable 5 and every other paid model on the leaderboard. A separate, independently maintained model catalog lists GLM-5.2 at $1.40 per million input tokens, confirming the price gap is real and not just a marketing number from the company that built it.

A different testing group called Kilo ran its own matchup days later, on a different task entirely, scoring how well each model planned a coding project rather than how well it built a website. This time Fable 5 narrowly won, 9.1 to 9.0.

Two scoreboards. Two winners. Same week.

There's an awkward postscript to both tests. The version of Fable 5 they used to set the bar is the same one that's been offline since June 12. Whatever you conclude about how close GLM-5.2 came to the frontier, the frontier model itself currently isn't something you can buy, route to, or build a product around. The open-weight challenger is the only one of the two still running.

About Shashi Bellamkonda

Loading bio…

LinkedIn Profile

Get new posts by email:

Subscribe

Powered by