# Auto: one model string for long coding-agent sessions

> Auto starts each coding-agent session on a fast, inexpensive model and moves to a stronger one when the work calls for it. At most one switch per session, so the prompt cache stays warm.

Source: https://modeless-web.vercel.app/auto

`model: "auto"`

Auto is made for long coding-agent sessions. It starts on a fast, inexpensive model and brings in a stronger one when the work calls for it.

## How it works

You usually can't tell from the prompt how hard a task is. Auto keeps reading the session while the agent works, and changes model when the task needs it.

1. **Start**: Every session starts on a fast, inexpensive model.
2. **Monitor**: Auto tracks project size, problem difficulty and task progress as the conversation goes on, until they cross the upgrade threshold.
3. **Upgrade**: The session moves to a stronger model and carries on from where it was.

Monitoring reads the task and the agent's work so far. See [what gets sent, and to whom](https://modeless-web.vercel.app/security).

**Good fit**: long coding-agent sessions (Claude Code, Codex, OpenCode, Cline); batches of tasks where cost per task matters; teams that don't want to re-evaluate models every month.

**Not a fit**: every request must go to the strongest model available (call that model directly); you need to pin one exact model version for compliance or reproducibility.

## Benchmarks

- **Auto**: 30% lower cost per task than Claude Opus 5.5 (SWE Pro, DeepSWE v1.1, Terminal-Bench 4.0)

### SWE Pro

Results coming soon.

Method: Agent: TBD · Tasks: TBD · Runs per configuration: TBD · Cost: measured tokens × public USD prices

### DeepSWE v1.1

Results coming soon.

Method: Agent: TBD · Tasks: TBD · Runs per configuration: TBD · Cost: measured tokens × public USD prices

### Terminal-Bench 4.0

Results coming soon.

Method: Agent: TBD · Tasks: TBD · Runs per configuration: TBD · Cost: measured tokens × public USD prices

## Model card

- Protocols: Chat Completions · Anthropic Messages · Responses
- Streaming: Yes, on all three
- Input: Text, images (routed to models that accept them)
- Tool calls: Yes
- Thinking parameters: Passed through; routed to models that accept them
- Context length: TBD
- Models used: A mix of fast open models and frontier models. The mix changes between versions; your console shows the model behind every request.
- Billing: each request at the price of the model that answered it, plus 5%.