A Route Around the Model
Anthropic’s Claude went down at the end of July, two days of elevated errors on the 29th and 30th, logged at status.anthropic.com, and in the week that followed the practitioner rooms I compare notes in filled with reports that the flagship model had been quietly downgraded and that scheduled jobs were misfiring, whilst one member found the same model noticeably better on his workloads. None of it could be proved in either direction, because none of the reports came with a benchmark. Dependence on one vendor hands you both, the outages and the ambiguity.
No operations manager would wire a plant to a single power line and call the matter settled; the line is somebody else’s asset, maintained and priced at somebody else’s discretion, so you keep a generator and you test it. Most organisations have wired their AI the opposite way, one vendor, one model, the working knowledge poured into that vendor’s proprietary formats, and no tested route to anywhere else.
Electricity at least fails loudly, but a model can degrade quietly, the output a shade lazier, the refusals a touch more frequent, and you have no way of knowing whether the change is in the model or in your own head. An operation that cannot tell whether its core capability got worse has already lost something, whatever the truth turns out to be.
The exposure is also larger than it was a year ago, because the work has moved. When AI was a chat window an outage cost you convenience, but once agents are wired into operations, running scheduled jobs overnight and acting in your name, an outage is a stoppage, and a quiet degradation is worse, because the work carries on, slightly wrong, at machine speed.
None of the three moves below requires a platform programme.
Keep a second route, and exercise it. If your work runs on Claude, keep a GPT or Gemini route warm, running a real workload through it on a schedule so the comparison exists before you need it. A fallback you have never run is a rumour.
Route by the shape of the work rather than the brand of the model. Code-shaped tasks and judgement-shaped tasks already favour different models, the ranking reshuffles with every release, and a team that can move work between routes treats the reshuffle as noise.
Own the substrate. Keep your instructions, your working knowledge and your standards in plain text that you control, behind a gateway that lets the same work point at a different model with a configuration change (LiteLLM if you want to run it yourself, OpenRouter if you do not). This is the rule we apply in our own builds and our clients’, i.e. standardise on the substrate, stay plural on everything above it. Our own delivery pipeline is wired this way, three model families behind one set of plain-text instructions, and plain text is the one format every model can read and write, so nothing about the arrangement is clever or expensive.
The objection is cost, and it is real, a second contract and the standing chore of exercising a route you mostly do not use. Price it against the week just gone, a two-day outage followed by five days of arguing about whether the model was still any good.
The test is cheap enough to run this week, so pick the one AI-dependent workflow that would hurt most if it stopped, and let’s run it once on your second-choice model, keeping the output for comparison. If there is no second choice and no way to move the work, you have found the real single point of failure, and it is not on Anthropic’s status page.
Mark Bunce is the founder of Trueform Consultancy, focused on redesigning how work gets done using modern technology.