Guide
Models and Ultrafast
Pick the model every agent in a swarm runs. Pass gpt-6.1-sol-ultrafast for GPT-6.1 Sol on OpenAI's Ultrafast tier.
From the CLI
Add
--modeltoswarm run. Needs CLI 0.0.9 or later; until it is out, use MCP below.acyclic swarm run --model gpt-6.1-sol-ultrafast \ --task "Fix the failing tests in tests/" --dir . --watch
From your coding agent (MCP)
The
spawn_swarmtool takes the samemodelargument. Ask your agent for it by name.spawn_swarm({ "task": "…", "model": "gpt-6.1-sol-ultrafast" })
Model ids
| model | What it runs | Per 1M tokens |
|---|---|---|
| (omit) | Platform default: GPT-5.6 Sol, with GPT-5.6 Luna for delegated children. | Per model |
| gpt-6.1-sol | GPT-6.1 Sol, standard tier. | $2 / $0.10 / $10 |
| gpt-6.1-sol-ultrafast | GPT-6.1 Sol on OpenAI's Ultrafast tier: the same model, served much faster. | $12 / $0.60 / $60 |
| claude-opus-5-5 | Claude Opus 5.5. Also claude-sonnet-5-5, claude-fable-5-1. | Anthropic list |
| gpt-5.6-sol | GPT-5.6 Sol. Also gpt-5.6-terra, gpt-5.6-luna. | OpenAI list |
Notes
- Prices are per million input / cached input / output tokens, billed at the provider's rate with no markup. Ultrafast is six times the standard rate, so set a budget.
- Every agent in the swarm uses the model you pass, children included; only the judge that picks between speculative attempts keeps the platform default. Choosing a model per role is only available on a dedicated control host.
- An id the host does not offer is refused before the swarm starts, and the error lists every id it does offer.
- A swarm that chooses a model runs on Acyclic's keys and is billed from your credit, even if your organisation added its own OpenAI key.
More: first swarm · swarms as an agent tool · CLI guide