A good model choice does not stay good on its own. Without governance, every team drifts back to the expensive default. The fix is policy that makes the cheap path the easy path, and a light review that keeps it that way.
A team can do the analysis, measure accuracy per dollar, pick the right model for each task, and ship a well tuned model mix, and then watch it decay over the next two quarters. New features get built on the most capable model because that is the fastest path to a working prototype. Engineers who were not part of the original analysis copy a snippet that points at the expensive model and never revisit it. The careful mix erodes one default at a time until the bill is back where it started. This is not a failure of skill, it is a failure of governance. Good model selection is not a project you finish, it is a standard you maintain, and maintaining it across many engineers and many services requires deliberate structure rather than goodwill. This is what that structure looks like.
The decay is structural, so understanding it points straight at the fix. The expensive model is the default for a reason: it is the safest choice during development, it forgives weak prompts, and choosing it requires no thought. The cheap model requires a decision, usually backed by a small evaluation, and decisions that require effort lose to defaults that require none, especially under deadline pressure. So unless the environment is arranged so that the cheap, correct choice is also the path of least resistance, entropy wins and everything drifts upward. Governance is the work of changing the defaults and the incentives so that the disciplined choice is the easy one, rather than relying on every engineer to fight entropy individually on every feature.
The foundation is a short written policy that says which model is the default for which kind of work. It does not need to be elaborate. It should state that simple, high volume tasks like classification and extraction default to the cheapest tier, that the workhorse tier is the default for general production tasks, and that the most capable tier is reserved for the specific task classes that have been shown to need it. Crucially, the policy should make the top tier the exception that requires a reason, not the default that requires no reason. When using the expensive model is a choice someone has to justify rather than the path of least resistance, the whole gravity of the system changes. A policy that fits on a page and inverts the default is worth more than a sophisticated framework nobody reads.
Policy alone does not hold; the tooling has to back it. The most effective governance move is to make the disciplined choice the default in code. That can mean a shared client library where the model is selected by task type rather than hard coded, a routing layer that applies the policy automatically, or templates and starter code that point at the workhorse model rather than the expensive one. When the easiest thing to do is also the right thing, drift stops, because there is nothing to drift toward. The opposite arrangement, where the policy lives in a document but the convenient code points at the expensive model, guarantees decay no matter how good the document is. Engineers follow the path of least resistance, so governance works by moving the path, not by exhorting people off it.
Some review is needed to catch what tooling and policy miss, but it has to be light or it will be ignored. The two moments that matter are the start and the periodic check. At the start, when a new feature or service is being designed, a quick model selection question in the design review catches expensive defaults before they ship and harden. Periodically, a review of where the spend is actually going surfaces the services that have drifted, so they can be corrected before they have run up a large bill for months. The aim is not to police every model call, which is unworkable, but to put a checkpoint where new defaults are set and another where existing ones are audited. Two well placed light reviews beat a heavy process that the team routes around.
Governance lands better when engineers can see the cost of their choices. A team that has no visibility into what each service spends on Claude cannot self correct, because the information that would prompt a better choice is invisible to the people making the choices. Surfacing cost per service, and ideally cost per request type, turns model selection from an abstract policy into a concrete number an engineer can act on. Most engineers, shown that a particular service is running a high volume cheap task on the expensive model, will fix it without being told, because the waste is now visible and the fix is theirs to make. Visibility does much of the enforcement that a heavier process would otherwise require, and it does it with the grain of how engineers actually work.
Model selection governance is the difference between optimization that holds and optimization that quietly unwinds, and that difference shows up directly in the number you commit to Anthropic. A buyer whose model mix is governed has a stable, defensible run rate, which means a commitment sized to reality rather than to drift, and a credible story of disciplined usage that strengthens every negotiation. A buyer whose discipline decays signs a commit against a low number and then climbs past it as the defaults creep back up, paying for the erosion twice, once in the higher bill and again in the overcommitment they sized wrong. Governance protects the savings you worked to find and protects the commercial position those savings built.
Governance is what makes routing, accuracy per dollar measurement, and the rest of the optimization program durable rather than temporary. It is the maintenance layer under the whole token optimization method. For the full playbook, including the policy template and the review cadence, read the pillar guide, the token optimization playbook. If you want governance designed and stood up across your teams, and the savings protected in your next Anthropic deal, get a quote and we will scope it with you.
We design model selection governance that holds, and protect the lower run rate in your Anthropic commit. Fixed fee or gainshare, no risk to you.
Get a QuoteWeekly intelligence on Anthropic pricing moves and the buyer side counters that work.