Pay for it, sheet 09 of 13
Sending each prompt to the right model, and surviving the deprecation clocks that retire the one you tuned for.
List every prompt pinned to a model with a shutdown date, and schedule the golden-set run before the date, not after.
← → move to the next sheet over, Esc zooms out.
Not every prompt needs the most expensive model. A router decides which model answers each request, by price, speed or how hard the task looks. The other half of the job is lifecycle: the model you tuned your prompts against has a retirement date, and its replacement behaves differently.
A router is the dispatcher at a cab company, sending the short hop to the small car and the airport run to the van. Lifecycle is the fleet: every car is retired on a schedule, and the replacement handles differently in the rain.
Routing by price and benchmark is now a commodity, with free options from cloud and gateway vendors. Routing by your own taste, by which model's work your team actually accepts, is still something you build.
Every link opened on 11 Nov 2026. The numbers match the squares on the map.
Reference, Anthropic
Which Claude models are active, deprecated or retired, with retirement dates and recommended replacements.
The catch. At least 60 days' notice. Put the page on a weekly check, because nothing polls it for you.
Product
Picks a model and a reasoning effort for each request, from the models your allow-list, data-retention rules and keys permit.
The catch. It optimizes for general quality and price. Your taste isn't in its training data.
Product, Microsoft
A deployable router that picks among models in Foundry for each prompt, with a balanced default mode.
The catch. The effective context window is limited by the smallest model in the pool.
Product, AWS
One endpoint that predicts response quality and routes between foundation models in the same model family.
The catch. Optimized for English prompts, and it can't learn from your own application's results.
Product
Model routing built for coding agents, aiming for frontier quality at a lower price.
The catch. The quality and savings figures are the vendor's own. Test it on your golden set.
Product
A gateway to hundreds of models through one API, with routing, failover to the next model, spend limits and observability.
The catch. One more party in the path of every prompt. Check where it stores what it sees.
Open source, Apache-2.0
A framework for serving and evaluating LLM routers, so you can train one on your own data.
The catch. Its repository hasn't been updated since August 2024. Read it as a worked example, not a maintained product.