Skip to main content
Multi-model is currently in beta. The core functionality is stable but the UI and configuration options may evolve.

​
Overview

Multi-model lets you assign different AI models to each phase of a workflow. Use an expensive reasoning model for planning, a fast model for implementation, and a completely different model for code review. This optimizes both cost and quality — you pay for reasoning only where it matters.
Credit Efficiency: Multi-model is a highly credit-efficient choice for routine tasks, using cheaper models for the bulk implementation phase while reserving premium reasoning models for the planning phase where they are needed most. Learn more in the Workflow Credit Efficiency reference guide.

​
When to Use

  • Cost-sensitive teams wanting to optimize credit spend
  • Complex tasks where planning quality matters more than implementation speed
  • When you want cross-model review (a different model catches different issues)
  • Teams that have benchmarked specific models for specific tasks

​
The Three Phases

​
How It Works

1

Select Multi-model workflow

Choose Multi-model when creating a task.
2

Configure models per phase

Assign agent presets to each phase. Create presets in Settings → Default agents with specific model and CLI configurations.
3

Execution

The task runs through each phase using the assigned model. Planning generates specs, implementation generates code, and review validates the output.
4

Review cross-model output

The review model provides an independent assessment. Since it has different assumptions than the builder, it catches different classes of issues.

​
Agent Presets

Agent presets are named configurations stored in Settings → Default agents. Each preset defines:
  • Agent CLI — which runtime (Zen CLI, Claude Code, Codex, Gemini)
  • Model — which model the agent uses
  • Configuration — execution mode, approval policy, tool permissions
Example presets for a multi-model pipeline:

​
Why Cross-Model Review

A model reviewing its own output tends to agree with its own reasoning. A different model brings different assumptions and catches different problems. For higher-stakes changes, running multiple reviewers in parallel gives broader coverage than a single reviewer.

​
Tips

Try the default configuration before tuning. It uses a reasoning model for planning and a fast model for implementation — a good baseline.
Within the review phase, you can spawn multiple subagents with different models to review the same changes simultaneously. See Subagents.
Opus and GPT-5.4 excel at planning and architecture. Gemini Flash and Sonnet are fast at code generation. Use the cheapest model that meets the quality bar for each phase.

All Workflows

Compare all available workflows

Multi-Agent Orchestration

Deep dive into agent presets and subagents

Supported Agents

Available CLI agents and models