Models and routing
Every model choice, with its reason.
Your policy decides which models your AI agents may use. Every run records the model it ran on and why.
mistral-small-latest·Light question, efficient modelEfficient
A short chat turn, so it ran on the provider’s efficient model.
Model routing decides which model answers a request. In Infinized it follows the AI agent's own setting and your workspace's routing policy, and every run records the route it took and why.
A routing policy per workspace decides which providers and models your AI agents may use. Every run records the model it ran on and the reason it landed there.
Fallback
When a provider fails, the run moves on.
Give an AI agent up to three fallback models. An outage is retried first; then the run moves to the next model on the list. Every attempt is recorded as its own run, with its reason.
- Set modelclaude-sonnet-5Provider outage · retried first
- Fallback 1mistral-medium-latestRan here · recorded as “Fallback model”
- Fallback 2gpt-5.4Next in line, not needed
On every run
Your keys. Your policy.
Infinized routes each run and writes the route down: the provider, the model, the tokens, the estimated cost and the reason. The key, the contract and the invoice stay with your provider, and Infinized adds nothing to the price of a token.
- Cost
- $0.0031
- Latency
- 1.2 s
- In
- 1,204
- Out
- 186
How it works
What the policy decides.
Why this model, on every run
Each run records its provider, model, tokens, estimated cost and the reason it ran there: the AI agent's setting, an economy turn, a flow step's tier, a reroute or your policy.
A routing policy per workspace
Allow the providers and models your AI agents may use, and set an expected cost per run that steers work away from the models above it.
Fallback when a provider fails
Give an AI agent up to three fallback models. A provider outage is retried first, then the run moves to the next model; a dead key or an empty quota does the same. A disconnected provider sends the AI agent to another one at the same tier. Every attempt is its own recorded run.
Light turns on a light model
A short chat turn runs on the provider's efficient model. Shadow mode records what every run would have cost on the lighter model, and the lighter route switches on once your evals pass the routing check.
A tier per flow step
A flow step can ask for a flagship, balanced or efficient model on the AI agent's own provider, so a routine step does not pay for the heaviest model.
Model calls kept in your region
Opt in, and the policy only calls providers inside your data residency, such as Mistral or OpenAI's EU endpoint for a European workspace.
Providers
The models you allow.
30 models from 3 providers in the catalogue, each with its tier. Your own OpenAI-compatible endpoint can join them.
Mistral AI
Open-weight models for chat, coding and embeddings.- mistral-large-latestFlagship
- mistral-medium-latestBalanced
- mistral-small-latestEfficient
- magistral-medium-latestReasoning
- magistral-small-latestReasoning
- ministral-8b-latestEfficient
- ministral-3b-latestEfficient
- codestral-latestCoding
OpenAI
GPT models, plus voice and embeddings.- gpt-5.6-solFlagship
- gpt-5.6-terraBalanced
- gpt-5.6-lunaEfficient
- gpt-6-astraFlagship
- gpt-6.1-solBalanced
- gpt-6-lunaEfficient
- gpt-5.4Balanced
- gpt-5.4-miniEfficient
- o3Reasoning
- o4-miniReasoning
- gpt-4.1Balanced
- gpt-4.1-miniEfficient
- gpt-4oFlagship
- gpt-4o-miniEfficient
Anthropic
Claude models for instruction-following and tool use.- claude-opus-5Flagship
- claude-fable-5Flagship
- claude-sonnet-5Balanced
- claude-haiku-4-5Efficient
- claude-opus-5-5Flagship
- claude-fable-5-1Flagship
- claude-sonnet-5-5Balanced
- claude-haiku-5-5Efficient
OpenAI-compatible. Your own models, on your own endpoint.
Where it stops
- Routing decides which model answers. The key, the contract and the invoice stay with your provider, and Infinized adds no markup to tokens.
Questions
What people ask first.
Read on in Learn
Which AI model should each AI agent use?Can we see why a run used a particular model?
Yes. Every run records the model it ran on and one of nine route reasons, from the AI agent's own setting to a reroute or your policy. The runs report over the API carries it too.
What happens when a provider has an outage?
The call is retried first. If the provider stays down, the run moves to the AI agent's next fallback model, and each attempt is recorded with its reason. A key that stops working is set aside and owners and admins are told.
Can we keep model calls inside the EU?
Yes, as an opt-in routing policy. With it on, providers outside your workspace's data residency are not called. Mistral and OpenAI's EU endpoint count as European.
See it at work in your organisation.
We start by listening to your question. Then we show you a demo that fits your organisation.