Introducing virtual models on AI Gateway. Use a single virtual model ID across agents to: • Centrally manage model strings and configs • Immediately update to new models • Enforce Zero Data Retention Read the blog post →

now i can break all my agents simultaneously with one click

finally an env var with a blog post

one gotcha before you repoint a virtual model to another lab: provider options are per provider, so an anthropic-only setting just stops applying on openai. create vmc/coding-agent-candidate on the new model, run evals there, then edit the real one

an alias that resolves on your side means the model can change under a caller without a deploy. log the resolved model id on every request, or the first sign of a swap is a batch of runs that suddenly read differently. what does the alias return when the target model is down?

beats find and replace every time Anthropic updates a date.

one central model name beats ten agents with hardcoded strings.

Model IDs hardcoded in every agent config was never going to last. A layer that resolves them makes sense.

Can I replay a failed run with the exact model config it used? Otherwise I might be debugging yesterday's failure with today's settings.

baking the model string into every agent is the real lock-in. a virtual id is just config you can flip

A single model ID across agents is a lovely way to retire the “which config still points at the old model?” scavenger hunt

gateway is 16x smaller than OR right? and they are the second biggest comp, no? nuts

nice, gives a layer to swap models and enforce retention policies

The strong part is precedence: virtual model settings win over request settings, so an agent can't disable ZDR or pick its own fallbacks. Model governance moves from code review into the gateway. The candidate-ID pattern for evals before switching production is a nice touch.

virtual model IDs solve the biggest operational nightmare of fleet-scale agent deployments: configuration coupling. 1. decoupling agent code from upstream lifecycle churn hardcoding specific model strings across dozens of agent services means every model deprecation or version bump requires 30 PRs and redeploys. routing through semantic virtual IDs (e.g. fast-triage, code-orchestrator) moves model selection from code to runtime configuration. 2. dynamic failover without session tear-down if an upstream provider has a regional outage or hits a 429 rate limit mid-task, an agent without a gateway simply crashes. virtual models enable instant circuit-breaking: transparently rerouting the request to an equivalent provider with zero state loss and zero client restarts. 3. centralized zero data retention (ZDR) policy enforcement you can't rely on individual developers or subagents to remember to pass proper privacy flags. enforcing ZDR and token budgets at the gateway boundary guarantees organizational compliance before requests ever leave your VPC. dynamic proxying is mandatory once you move beyond toy scripts to 24/7 autonomous workloads

what if my users don’t want to pay api prices

Model aliases fix the most boring real problem in AI teams: someone's prod still pinned to last year's model.

The useful question is governance, not convenience: can a virtual model be pinned and rolled out gradually, or does 'immediately update' mean every agent silently switches at once? One ID is great until a model change alters behavior with no code diff to blame.

One virtual model ID across agents is a clean abstraction. Swap models without rewriting code. That is how agent systems stay flexible.

One model ID to manage configs, switch models, and enforce zero data retention makes running multi-agent systems a lot simpler.

for the zero data retention part: if the model a virtual id points at has no zdr route, does the request fail outright or quietly get served somewhere else?

@bot take a look at this

The upgrade checklist would like to apply for early retirement.