The Frontier Is on Loan

On Wednesday, Fable 5 was the most capable model you could buy. By Friday it was gone.

On 12 June, the US government directed Anthropic to suspend all access to Fable 5 and its restricted sibling Mythos 5, worldwide, for every user, citing export-control and national security concerns over a potential jailbreaking method. Anthropic disputed the call, saying the technique exposed only minor, already-known weaknesses found in competing models too. The dispute will run its course. The fact on the ground did not wait: the top of the range was available one day and unreachable the next, for reasons no customer had any say in.

We published a piece on Fable 5 the week it landed. The argument was that the ceiling moved and the job did not, that for a solo operator the work is still to point capable models at the right problems and supply the judgment they cannot. The suspension is that argument playing out in public, faster than we expected. The frontier disappeared overnight, and for anyone whose daily work ran on the dependable tier below it, whether Anthropic's Opus and Sonnet or another provider's models entirely, the day's work did not change at all.

You could watch the adjustment happen in real time. Within a day of the suspension, a steady stream of developers who had been running Fable for agentic coding moved back to OpenAI's Codex, which has become the common destination for anyone leaving a Claude-based setup, helped by OpenAI's edge on terminal-native coding. Others reached for Google's Gemini command-line tools or open harnesses such as Cline. What stood out was the speed. The work moved because it was portable: the harness, the prompts and the judgment carried across, and only the engine underneath changed.

Here is the part worth keeping. The frontier is on loan. Whatever sits at the very top of the range is the most exposed thing in your stack: newest, most scrutinised, most likely to be repriced, rate-limited, rerouted through a safety classifier, or, as we have now seen, pulled entirely. You do not control any of that. You rent the frontier on terms someone else sets and can change without notice.

The dependable tier is different, and it does not belong to one company. Anthropic's Opus and Sonnet, OpenAI's GPT-5.5, Google's Gemini, DeepSeek's V4: any of these will do the overwhelming majority of real work for a small business, and they are stable, broadly available and priced for daily use. They are as close to "yours" as a model gets, and, crucially, there are several of them from different providers. The field is wider still, with xAI's Grok and Mistral's open-weight models among the alternatives, so a small team always has somewhere to go. If your operation runs on that tier, a single frontier model vanishing is a headline, not a crisis. If you had welded every workflow to the top of one vendor's range to feel ahead, you would have spent Friday morning rebuilding.

That breadth is the real insurance. The operators who barely noticed this week were not the ones who had picked the best model. They were the ones not tied to any single one. Keep your prompts, your context and your workflow in a form you can point at a different model, and the question "what happens if my provider's strongest model disappears" has a dull, welcome answer: you point at the next one and carry on.

There is a floor beneath even switching providers. Open-weight models, from DeepSeek and Qwen to Mistral and Meta's Llama, can be downloaded and run on hardware you own, which makes them the one option no government order and no pricing decision can reach. The caveats are honest: the open models still trail the frontier, and a machine that runs a capable one well is expensive to buy and to power, so it is slower and a step behind what you rent from the top labs. But viable does not mean equal. This is where the difference between a prepared business and an exposed one shows. If you had welded everything to one hosted model and it went away, the operation could stall. If you had a self-hosted open model ready to take the load, you would drop to something rougher and slower and keep trading, entirely on your own terms. Most people will never need to stand on that floor. Knowing it is there, and having it prepared, is what changes how exposed you really are.

So the question for a solo operator was never "what is the best model available". It is "what can I depend on". Those are different questions, and the gap between them is exactly the risk that showed up this week. Depend on the tier you can count on. Reach up to the frontier for the occasional job that genuinely earns it, the long migration or the multi-day research run, and treat that reach as a bonus you can lose, not a floor you stand on.

None of this is an argument against capability climbing. It will keep climbing, and the next frontier model will arrive soon enough. It is an argument about where you put your weight. Build on the branch that holds. Admire the new growth at the top, use it when it is there, and do not be standing on it when it goes.

The frontier is on loan. The workhorse is yours. The job, still, is the same.

Three things to do with this. Run your daily work on the dependable tier and prove to yourself it is enough, because it almost certainly is. Keep your setup portable, so the model underneath is a choice you can change in an afternoon rather than a foundation you have poured, with any frontier-dependent task kept isolated, a losable option and not a load-bearing wall. And notice, the next time a launch tempts you to rebuild everything around it, that the most powerful model in the room is also the one you hold on the shortest lease.