Models

Every task runs on a coding agent and a model. In the app both come from the agent you assign the work to, so there is no per-task model choice. Over the API the choice is per session, so two sessions under the same task can run different models.

Choosing a model

The model belongs to the agent. Open the agent from Agents, go to its Settings, and select Model. The picker groups models under the coding agent that runs them:

  • Claude Code (Anthropic)
  • Codex (OpenAI)

Pick a coding agent, then a model within it. Leave it on Default to follow the account, which is Opus 4.8 under Claude Code. The choice applies to the tasks that agent starts from then on; a session already running keeps the model it started with. To run the same work on a different model, assign it to an agent set to that model.

Claude Code (Anthropic)

ModelNotes
Opus 4.8Default.
Opus 5Current Opus release. Runs against a one-million-token context window.
Fable 5Most capable. Costs more per session, so pick it explicitly.
Opus 4.7Prior Opus release.
Sonnet 5Faster and lower cost than Opus. Runs against a one-million-token context window.
Sonnet 4.6
Haiku 4.5Fastest and lowest cost.

Context Window

A session on Opus 5 or Sonnet 5 works against the model's full one-million-token window. The other Claude models work against 200,000 tokens, and a session that outgrows that window cannot take another turn.

The difference shows on long-running work: a task that reads a lot of a large repository, or that runs for hours across many turns, keeps going on Opus 5 or Sonnet 5 where it would otherwise stop.

Codex (OpenAI)

ModelNotes
GPT-5.6 SolFrontier tier.
GPT-5.6 TerraBalanced intelligence and cost.
GPT-5.6 LunaEfficient, for high-volume work.
GPT-5.5Prior flagship.
GPT-5.4The generation before that.

An unknown or unavailable model falls back to the agent's default rather than failing the task, so the session always runs on a supported model.

Choosing a model over the API

Set the agent and model when you create a task or session with the harness and model fields. See Tasks and Sessions.