Until now, every agent in a LoopLlama crew ran on a Claude model. That was a fine place to start — one provider, one set of behaviors to learn — but it turned model choice into a platform decision instead of a per-step one. Starting today, a workflow can route each step to any model from any major provider, in the same run, with the same tools, approvals, and checkpoints. And we've changed how model usage is billed: at the provider's published list price, passed straight through, with zero markup.
Eight providers, one workflow definition#
A workflow still has a default model, and any agent can still override it. What's new is what you can put in that field: Anthropic Claude, OpenAI GPT, Google Gemini, xAI Grok, DeepSeek, Mistral, Perplexity Sonar, or any OpenAI-compatible endpoint you host. Bare ids like gpt-5.4 or gemini-2.5-flash are routed by their naming convention; prefixed ids like openai/gpt-5.4 or custom/my-model make the provider explicit. Unroutable ids are rejected with a 400 when you save the workflow, so a typo never reaches a run.
- Tool calling, connector capabilities, and approval gates behave identically on every provider.
- Human-in-the-loop pauses and resumable checkpoints work the same way regardless of which model paused.
- Per-step traces record the provider's reported token usage and the step's cost, so a mixed-provider run reads like a single stack trace.
- GET /api/models lists what the platform can route to right now, with list prices.
Live web research, with citations#
Perplexity's Sonar models are the first web-grounded models on the platform. A Sonar step searches the live web on every call and returns cited sources, which makes it a natural research or fact-check stage in front of a drafting step on a frontier model. Sonar, Sonar Pro, Sonar Reasoning Pro, and Sonar Deep Research are all available; Deep Research runs long and bills reasoning and citation tokens separately, and we pass that through exactly as reported. Sonar models don't call tools, so pair them with a tool-capable model for anything that has to touch your systems.
What models cost now#
Your plan covers the platform: the workflow steps your agents run, the connectors, the traces, the approvals. Model usage is a second, transparent line. Each step records the provider's input, output, and cached tokens — plus any per-request search fee — priced at that provider's list rate, and the sum is invoiced to paid plans each period with no markup. The usage dashboard shows the cost per step, per run, and per model so you can see exactly where the tokens are going. On the free Developer plan, model usage is covered by your included steps.
An unlisted model — a brand-new release that isn't in the catalog yet — is billed at its provider's highest listed rate until we add it with its real price, so it can never be under-billed. The full catalog, with current list prices, is on the pricing page and in the docs.
How teams are using it#
The pattern we see most is the one we built for: a cheap, fast model to classify or extract, a Sonar step to research with sources, a frontier model to reason or draft, and a different frontier model to review. Four providers in one run, no glue code, and the review step costs what the review step costs. Set a default model on the workflow, override the steps that need something else, and watch the per-step traces to tune from there.