Models

Every agent runs on one model, named in its config: claude.model for Claude Code agents, codex.model for Codex agents. Omit the field and the agent runs the default (Claude Opus 5 for Claude Code agents). Ellipsis routes every model call to whichever provider setup your account uses; the agent's behavior is identical either way.

How model routing works

Four setups, from zero configuration to full control:

• Ellipsis-managed (the default): agents run on Ellipsis's own model access. Nothing to set up; tokens are metered to your Ellipsis subscription.
• Your Anthropic key: connect an Anthropic API key and tokens bill your Anthropic account, against your credits and commitments.
• Your AWS Bedrock account: connect AWS credentials and tokens bill your AWS account.
• Your own LLM proxy: model calls leave through a gateway you run, for compliance or logging.

The key and Bedrock setups are account-level: connect one and it routes every session, and at most one of the two routes at a time (verifying one disconnects the other). The proxy is different: register one or more gateways, then opt individual agents in with the llm.proxy config field. Registering a proxy moves no traffic by itself.

Routing changes who pays for tokens, not how spend is controlled. Session cost breakdowns always show real token spend, so spend limits bind whichever provider the tokens run on.

Choosing a model

Model choice is a per-agent cost and capability call, and the right answer follows from the job:

• Digest and summary jobs run cheap models. Splitshift's Recent work summary reads the week's merged pull requests and writes a 150-word summary on Claude Haiku 4.5, at a fraction of Opus cost. Retrieval and summarization rarely reward a frontier model.
• Judgment jobs earn the frontier model. Splitshift's skill maintainer decides whether a diff contradicts a written claim, and a cheap model fails that job in both directions: it rewrites skills that were fine, or misses the rename that matters. How to keep your skills from going stale shows why paying for the stronger model is the point there.
• When unsure, start with the default. New agents run Claude Opus 5 unless the config says otherwise. Move an agent down to Sonnet or Haiku once its sessions show the job does not need it.

A wrong model choice costs money, not safety: budgets cap what any session or agent can spend regardless of model. The full model list with rates is at Ellipsis-managed models.