The version number is the trust contract, and routing breaks it

In April, a lot of developers became convinced that Claude had gotten dumber. Work that held together in February stopped holding. The assistant lost track of what it had done twenty minutes earlier, ignored the project instructions sitting in the repository, and stopped updating its own memory files. A GitHub issue thread on Anthropic’s Claude Code repository filled up with people comparing symptoms. Explainer posts appeared with titles asking why Claude was worse now.

The behavior was real. The explanation everyone reached for was wrong, and how wrong it was is the useful part.

What actually happened, for anyone who spent the spring shipping instead of reading changelogs: a large AI model can be run at different levels of effort. Same model, same weights, but you can tell it to think harder or to answer faster, and answering faster is cheaper and feels snappier. On March 4, 2026, Anthropic turned that dial down inside Claude Code, its coding tool, from high effort to medium. The reason was latency. Users wanted faster responses. The model itself was untouched, and so was the API. Two other product-level changes landed in the same window: a caching bug that dropped context, and a cap on how much the assistant would write.

The complaints started around April 15. The fix shipped April 20. Anthropic published a post-mortem naming all three causes on April 23. Count from the dial turning to the public explanation and you get roughly seven weeks, during which users had no way to attribute the change to anything except the model’s intelligence.

This matters well beyond one vendor’s bad quarter, because routing is the largest cost lever in AI products right now and the case for it is genuinely strong. The Pragmatic Engineer, a software engineering newsletter, reported in July that the gap between the cheapest usable model and the most capable one runs to roughly 100×, and that teams cut 40 to 85% of their spend by sending easy requests to cheap models and hard ones to expensive ones. Martin Fowler reported the same month that one enterprise’s annualized AI bill went from $5 million to more than $120 million inside a year. At those numbers routing stops being an optimization and starts being solvency.

The pitch carries a clause almost everyone reads past. Those savings arrive, teams report, “without visible quality degradation.” The only cost anyone itemizes is an engineering one: routing is hard to instrument, so when something goes wrong you struggle to reconstruct which request went where. An observability problem. Something for the platform team’s backlog.

Anthropic just ran that experiment in public, and the invisible cost showed up in a currency the ROI model has no row for: seven weeks of users doubting a capability that never changed.

Take the caveat first, because it sharpens the point rather than softening it. This was an effort setting inside one model, not routing between models, and three changes shipped into the same window, so you can’t pin the whole regression on the dial.

But look at what stayed still. Every study of routing quality confounds the tier with the weights: you moved to a smaller model, so of course the answers changed. This episode didn’t move models. It said Opus 4.6 on March 3 and Opus 4.6 on March 5, and Anthropic confirmed the weights and API were untouched throughout. Whatever the split between the three changes, all of the observed damage came out of the product-configuration layer that routing economics treats as free. The layer everyone calls invisible was the only layer that moved.

Notice what the version number did through all of this. Nothing.

That’s the part I’d put in front of anyone about to ship a router. The version number is the trust contract your users actually read. It’s the only handle they have on “is this the same thing I evaluated last month.” Effort settings, routing rules, cache behavior, and verbosity caps all sit underneath it and change without a version bump. Naming your tiers doesn’t fix that, and the governance story around named tiers has it backwards. A label that holds still while the thing beneath it drifts is worse than no label. It reads clean, so nobody looks. The control point is real for whoever declares the policy and worthless for whoever consumes the output.

Which exposes the assumption sitting under the whole routing case. Routing economics is written from the operator’s chair, where any trust cost stays in-house and your own engineers grumble at your own tool. Route inside a product and the person pulling the lever and the person eating the consequence are different people. You spend their trust to buy your latency and your margin, and they hold no instrument that would tell them you did it.

Trust in a long-running agent works like a switch. The entire value of a multi-step agent is that you stop checking its work. Start checking again and the token savings are beside the point, because you’ve put back the human cost the automation existed to remove. Shaving a few percent of quality to save a large percent of cost assumes those two percentages trade against each other. You can’t convert one into the other.

Anthropic appears to have reached the same conclusion. Claude Sonnet 5, released June 30, ships with selectable reasoning effort. The dial moved from the vendor’s side of the wall to the customer’s, which reads to me like a concession.

So the position I’d argue for, and hold loosely enough to be argued out of: if a configuration change alters the behavior users attribute to the model, it belongs inside the version contract. An effort-tier change or a routing change should force a version bump the way a weight change does. That’s expensive and nobody selling a router wants it. It’s also the only thing that converts a silent product-layer change from an undisclosed liability into a disclosure.

The honest gap is that I can’t price the other side yet. Nobody published adoption or retention data from that seven-week window, so the size of the trust bill remains unknown, plausibly including to Anthropic. Route anyway, probably. The pressure is real and so are the savings. Just go in knowing the number in circulation prices exactly one side of the trade, and the other side came due in public, for the vendor best positioned to have seen it coming.