Choose a model for a workflow, then verify what answered

Decide what the choice is for

Start from the trade-off, not the model name. GitHub Copilot's changelog for 14 September 2026 says auto model selection now offers three tiers - efficiency, balance and intelligence - and describes the tier you choose as how you want auto to weigh cost, quality and response time. If your client has a setting like that, set it deliberately: it decides the priority your default now follows.

If the workflow is reproducible, the safer habit is still to name the model and version you tested rather than leaving the choice to an automatic rule.

Know whether a name is pinned or follows the family

Some catalogue entries are designed to move. OpenRouter lists an entry called OpenAI: GPT Sol Latest which the listing says always redirects to the latest model in the OpenAI GPT Sol family; the entry page shows the model it currently points at, and at the time of writing it displays Latest GPT-5.6 Sol. A fixed version name stays where you set it.

So there are two different jobs. An alias is convenient for a prototype or a task where you always want the newest model. A pinned version is what you want when a result has to be reproducible or when a change of model would move your costs.

The catalogue list itself is public - OpenRouter's models endpoint returns the entries and their ids - so you can check what an alias resolves to before you ship, and again after any upgrade.

Make the cost numbers agree

Two documented settings address the gap between what a client reports and what you are billed. Claude Code v2.1.268 added a gateway pricing setting: with pricing: set in gateway.yaml, the release notes say signed-in clients receive the same rates through managed settings, so /cost and telemetry match the spend meter.

Claude Code v2.1.267 added maxEffortLevel, which the notes describe as a top-level or per-model setting under modelSettings that caps the effort level on every provider, including Bedrock, Vertex and Foundry, while still allowing a lower level to be picked. A cap is the simplest way to stop a single step from spending more than the task is worth.

Bound what a step may do

If the model can run commands, limit the reach of each one. Claude Code v2.1.271 added per-command allowed_domains for Bash, PowerShell and Monitor in auto mode with sandboxing: the hosts a command needs are reviewed with it and opened for it alone, and other hosts are refused.

This is the same idea as an effort cap, applied to the network instead of the budget: give the step what it needs, and make everything else fail loudly rather than silently succeed.

Re-check after each upgrade

A model change usually arrives inside an upgrade you already ran, so re-check at a fixed point: re-read the catalogue entry for anything that follows a family, read the release notes for the version you installed, and use the place your client reports which model answered. Our guide Set a Copilot auto-model tier, then check which model answered covers where Copilot shows that.

Keep the check short and repeatable - our guide Check for AI changes once a week with four public endpoints is built for exactly this, and two of the pages linked below are the kind of endpoint it points at.