Choose a model
See which models Kortecx offers your Apps today, and choose the one that answers.
Today, Apps run on Kortecx hosted models, or on a model provider you connect with your own key.
What the choice gives you
- You decide what "the model" means for your runs. Kortecx can offer more than one, and the one that answers is a choice rather than a fixed property of the product.
- It is a setting, not a one-way door. Point your workspace's runs at Kortecx hosted models or at a vendor you connect, and change it back; what you author on top of it does not change, unless an app pins a model of its own.
Choosing which model answers
Open Models to see every model this runtime offers and which one answers when a run does not name one for itself. That is the screen to check before you author a step against a particular model, because a step that names one the runtime is not serving is refused.
Loading or pulling a model is not offered on the managed service.
Use a Claude model on your own connection
When your workspace has an Anthropic connection, a run can name the Claude model it runs on. The connection supplies the key, and the run's cost is an estimate at Anthropic's list price, read on a stated day.
Add the connection
Open Models, add an Anthropic provider with your key, and press Test key. The models the key can reach are listed with the provider.
Pick the model for a run
When you run a task, pick that provider and the model.
| Model | Good for | List price per million tokens (in / out) |
|---|---|---|
| Claude Opus 5.5 | Hard, multi-step work: planning, code, long tool use | $4 / $20 |
| Claude Sonnet 5.5 | Everyday agent and enterprise work, faster | $2 / $10 |
| Claude Haiku 4.5 | Short, cheap steps: extraction, routing, summaries | $1 / $5 |
The prices are the ones dev served on 2026-10-10, each read from Anthropic's price list on a stated day. A model with no price row still runs, but its run has no cost.
What Kortecx sends each model is decided per model:
- Opus 5.5 and Sonnet 5.5 always think before they answer, at each model's default effort level.
- If a model declines a request, the run fails and says so. It never silently moves to another model.
- If the provider refuses the request itself, the run fails and says that instead.
What naming a model actually does
Naming one is a constraint the runtime checks before anything runs. A step that names a model the runtime is not serving is refused at authoring time — you find out while you are writing the workflow, not three steps into a run that has already sent an email. Leave the name out and the runtime binds the model it is serving for you. Steps that do no reasoning need no model at all.
See Workflows for where a model step sits in a chain.
Limits
Where a runtime is not serving a model at all, the commands that need one say so rather than returning something you might read as an answer. Features built on top of a model — grounded search, durable memory, an agentic turn — are unavailable in the same visible way rather than quietly degraded.
Read what a run cost
What a run cost — an estimate at the model vendor's dated list price for runs on your own key, with a daily cap — not a bill.
Guarantees
Four promises you can check: work resumes instead of repeating, an irreversible step can wait for you, an ungranted action is refused, and every run can be read.