The right model for each step.
Two makers, one platform
Our agents use OpenAI's GPT models and Anthropic's Claude models, both through Microsoft Azure AI Foundry under Microsoft's enterprise terms. Prompts and outputs are not used by us or by Microsoft to train models.
You do not pick a model or hold a contract with either maker. We assign a model to every step of every agent, and keep each assignment under review.
How the work splits
We group models by what they cost to run, at the maker's list price per million output tokens: small up to $5, mid-size up to $15, frontier above that. Most steps in a workflow are reading, sorting and extracting. Where a small model gives the same result in a side-by-side comparison, it does the work, faster and for a fraction of the cost. Frontier models are kept for the steps that need them.
Production figures for the 30 days to 3 October 2026: 52,071 agent runs across 19 live workspaces, our own included.
Small
Reading, sorting and extracting at volume.
Share of runs94.7%
Share of tokens64.8%
Mid-size
Summaries, scoring and checking.
Share of runs4.9%
Share of tokens32.1%
Frontier
Long-form drafting and the hardest reasoning.
Share of runs0.5%
Share of tokens3.1%
Running in production
Every model that ran an agent in production in the last 30 days.
| Model | Made by | Tier | Share of runs | Share of tokens |
|---|---|---|---|---|
| GPT-5.6 Luna | OpenAI | Small | 52.3% | 26.0% |
| GPT-4.1 mini | OpenAI | Small | 30.6% | 30.5% |
| GPT-4o mini | OpenAI | Small | 9.3% | 3.9% |
| GPT-4.1 | OpenAI | Mid-size | 4.8% | 31.6% |
| GPT-6 Luna | OpenAI | Small | 2.3% | 4.3% |
| GPT-5.5 | OpenAI | Frontier | 0.2% | 1.8% |
| GPT-5.6 Sol | OpenAI | Frontier | 0.2% | 0.2% |
| Claude Opus 5 | Anthropic | Frontier | 0.1% | 1.1% |
| GPT-5.4 mini | OpenAI | Small | under 0.1% | 0.2% |
| Claude Sonnet 5 | Anthropic | Mid-size | under 0.1% | 0.4% |
| GPT-6 Sol | OpenAI | Mid-size | under 0.1% | 0.1% |
| GPT-4.1 nano | OpenAI | Small | under 0.1% | under 0.1% |
Recently deployed, not yet in production work
Added to our production platform in the last three months. An agent moves to one of these only after the comparison described below.
| Model | Made by | Tier | Deployed |
|---|---|---|---|
| Claude Sonnet 5.5 | Anthropic | Mid-size | 29 September 2026 |
| Claude Opus 5.5 | Anthropic | Frontier | 22 September 2026 |
| GPT-6 Astra | OpenAI | Frontier | 6 September 2026 |
| Claude Opus 4.8 | Anthropic | Frontier | 31 August 2026 |
| GPT-5.6 Terra | OpenAI | Mid-size | 31 August 2026 |
How a model change is decided
A new model is not a reason to switch. When one reaches our platform:
- We replay real work from an agent on its current model and on the candidate, and score both against that agent's own checks.
- A person reads the comparison and decides.
- The change goes to production through a reviewed release that records who approved it and why.
Every week we also review each agent that ran: what it cost, how long it took, how often it failed or ran out of room for its answer, and whether its model is due to be retired.
Model changes and what decided them are posted in News.