Compare models

Open Compare on your Mac or in the web app. Specs compares published model information. You can also share a public comparison: its model selections are saved in the address.

In Tasks, choose a Git project and two to four exact models. Pick each provider account, reasoning effort and supported speed setting. Auto model selection is unavailable. macOS uses eligible connected providers, including local models. Cloud comparisons support Claude and tool-capable OpenRouter models, subject to your existing Pro and rollout access.

Every lane starts from one pinned committed revision in a separate Git worktree. Uncommitted source changes are excluded; acknowledge this before running. Worktrees isolate checkouts, while each attempt keeps the existing permissions.

Read each lane’s streaming transcript, approvals, questions, changes and recorded execution evidence. Stop interrupts one lane; Stop all interrupts every lane. Retry reissues the latest task in the same conversation and worktree, keeping earlier evidence. Retry or explicitly exclude failed lanes before sending a shared follow-up. Every included lane receives identical text and keeps its own context.

Model settings are fixed after launch. Start a new comparison to change them. Continue records your preferred result and opens its existing chat; it does not apply or merge changes into the source checkout.

Ask an AI judge is optional. Choose its model and account explicitly. The judge reviews recorded evidence in read-only mode and makes a recommendation; it cannot apply changes.

Published benchmark results retain the model variant, benchmark version, harness, settings, source and evaluation date when available. Missing information stays unknown. Benchmark snapshots are reviewed repository data, separate from your Codz task measurements. Provider charges, estimates and unavailable costs are distinct; subscription usage is not a per-task bill.

Mac history is local and web history is cloud history. Removing a comparison from history removes its coordination record while preserving child chats and worktrees.