# How AutoRouter picks a model for each turn Authors: AutoRouter Published: 2026-10-23T02:20:01.884Z Reading time: 8 min Canonical: https://autorouter.com/blog/how-autorouter-picks-a-model Publisher: AutoRouter, the Mac app for Codex and Claude Code (https://autorouter.com) > AutoRouter, the Mac app for Codex and Claude Code, picks a model for each turn in order: your rules first, then how hard the turn is, then only the models that clear the quality bar for it. Hard work goes to the most capable model your ChatGPT or Claude plans cover, and to credits only when none is strong enough. Everything else goes to the model that costs you least, which is often your plan or a low-cost open model. [AutoRouter, the Mac app for Codex and Claude Code](https://autorouter.com), picks a model for each turn in this order: your rules first, then how hard the turn is, then only the models that clear the quality bar for that turn. Hard work goes to the most capable model your ChatGPT or Claude plans cover, and to credits only when none is strong enough. Everything else goes to the model that costs you least, which is often your plan or a low-cost open model. AutoRouter adds 20+ models to Codex and Claude Code, plus one more choice: AutoRouter, which picks the model for each turn. You choose it from the model menu in Codex or with `/model` in Claude Code, and every turn after that shows its model, what it cost and why. AutoRouter wrote this post, and it describes routing as of October 22, 2026. ## In what order does AutoRouter pick a model? AutoRouter checks your rules first, then difficulty, then the quality bar, then your plans, and cost last. 1. **Your rules.** If one of your rules applies, it decides the turn. 2. **Difficulty.** Otherwise AutoRouter works out how hard the turn is. 3. **Quality bar.** Only models that clear the quality bar for that turn stay in the running. 4. **Your plans.** Hard work goes to the most capable model your plans cover, and to credits only when none is strong enough. 5. **Cost.** Everything else goes to the model that costs you least, which is often your plan or a low-cost open model. When a task is unclear, AutoRouter sends it to a model that can do it well and never to the cheapest one. ## What do your rules decide? A rule that applies to a turn decides that turn, before AutoRouter weighs difficulty, quality or price. You write rules in plain English, such as one that sends UI prompts to frontier models and another that sends git commands to a low-cost model. These are the example rules on AutoRouter's landing page: - "Use Claude Opus 5.5 for UI work." - "Prefer GPT-6.1 Sol for code reviews." Your rules come before AutoRouter's recommended ones, and you can see how often each rule is used. Every model has to clear the quality bar for a task unless your own rule names it, so that choice stays with you. To send every turn to one model, pick it in the model picker instead: turns then go straight to that model, and AutoRouter routes only while it is the model selected. ## What are AutoRouter's built-in rules? AutoRouter has six built-in rules, and your own rules come before them: - Hard work goes to your strongest model. - Quick tasks go to the cheapest capable model. - Unclear work is never treated as a quick task. - A task that continues keeps its model, or moves up. - Your plans come first for hard work. - Untested models are never picked on their own. ## How do difficulty and the quality bar narrow the choice? AutoRouter works out how hard each turn is, then considers only the models that clear the quality bar for it, and price counts only after a model clears that bar. Hard work goes to the strongest model your plans cover, even when a cheaper one comes close. AutoRouter sometimes misjudges how hard a task is, and you can rate the pick when it does. AutoRouter doesn't publish benchmarks yet. Whether sending some turns to cheaper models lowers the quality of your results is covered in [Does model routing make your results worse?](/blog/does-model-routing-hurt-quality) ## Which model gets a refactor, a commit message or a failing test? In the examples on AutoRouter's landing page, a six-file refactor goes to Claude Opus 5.5 on the Claude plan, a commit message goes to a low-cost open model, a test that only fails on CI goes to GPT-6 Astra on the ChatGPT plan, and UI work goes to Claude Opus 5.5 because a rule says so. These four turns are illustrations; real routes follow your plans and rules. | Turn | Where it goes | Why | | ----------------------------------------------------------------- | ---------------------------------------------- | -------------------------------------------------------------------------------- | | Refactor: a change across six files | Claude Opus 5.5 on your Claude plan | Hard work goes to your strongest model, and your plans come first for hard work | | Commit: write the commit message | A low-cost open model; your plan isn't touched | A commit message is quick work, and quick tasks go to the cheapest capable model | | Debug: a test that only fails on CI | GPT-6 Astra on your ChatGPT plan | The task is hard, so it goes to the strongest model your plans cover | | Design: UI work, with the rule "Use Claude Opus 5.5 for UI work." | Claude Opus 5.5 | Your rule applies, so it decides the turn | A hard turn that no model on your plans is strong enough for goes to a model outside your plans, paid from credits. ## Why does the model for each turn matter? The models AutoRouter chooses between are priced very differently: as of October 22, 2026, GPT-6 Astra lists at 100 times GPT-6 Luna's price per token on OpenAI's API. These are standard-tier API list prices in US dollars per million tokens, as of October 22, 2026, from [Anthropic's pricing](https://platform.claude.com/docs/en/about-claude/pricing) and [OpenAI's pricing](https://developers.openai.com/api/docs/pricing): | Model | Input, per million tokens | Output, per million tokens | | ----------------------------------------------- | ------------------------- | -------------------------- | | GPT-6 Astra | $10 | $50 | | Claude Opus 5.5 | $4 | $20 | | Claude Sonnet 5.5 | $2 | $10 | | GPT-6.1 Sol | $2 | $10 | | Claude Haiku 5.5 (prompts up to 100,000 tokens) | $0.10 | $0.50 | | GPT-6 Luna | $0.10 | $0.50 | [Anthropic's model guide](https://platform.claude.com/docs/en/about-claude/models/choosing-a-model) points complex agentic coding, such as large-scale refactoring, at Claude Opus 5.5, and real-time, high-volume and subagent work at Claude Haiku 5.5. It also describes pairing a lower-cost model with a frontier model so that most tokens are billed at the lower rate. [RouteLLM](https://arxiv.org/abs/2406.18665), a 2024 paper by Ong and colleagues, trained routers that send each query to a stronger or a weaker model and cut cost by more than 2 times in some cases without lowering response quality. That is the paper's result on its own benchmarks, and it is not a measurement of AutoRouter. Turns on your ChatGPT or Claude plan run on the plan, so these API prices don't apply to them. Paid models through AutoRouter are billed at the provider's price plus 10%, from AutoRouter Pro's included usage ($20 a month) first and then from credits you top up. AutoRouter's landing page works through an example week, which is an illustration and not a promise: 30% hard, 45% medium and 25% quick turns, at Claude Opus, Sonnet and Haiku 5.5 list prices with cached input at a tenth of input. Pro sends 85% of the quick turns and 75% of the medium turns to open models, which its included usage pays for. In that example week, as the landing showed it on October 22, 2026, AutoRouter's Free tier does 1.9 times the work of using a frontier model for everything, and AutoRouter Pro does 2.9 times. ## How can you see and steer every pick? Every turn shows which model handled it, what it cost and why, and you can steer picks in three ways: - **Rate a pick** that falls short. The AutoRouter team reviews real routing decisions and every rating people send, and ratings retune the router. - **Write a rule** in plain English for work you want handled a certain way. - **Pick a model yourself** from the model menu in Codex or with `/model` in Claude Code. Turns go straight to that model until you choose AutoRouter again. You can also see which models do best on which tasks before you set a rule or pick a model. The team tests every new model as it becomes available, and new models show up in the picker right away. Benchmarks and automatic evaluations on real coding tasks, and routing that learns on its own, are coming soon and aren't available yet. ## Questions ### What happens when my ChatGPT or Claude plan runs out? When a ChatGPT or Claude plan hits its limit, AutoRouter Pro keeps those turns going from its monthly included usage until the plan resets. If the included usage runs out too, AutoRouter asks you to top up. [How Claude Code usage limits work](/blog/claude-code-usage-limits) explains the plan limits themselves. ### Does AutoRouter earn more when it picks a paid model? Yes, AutoRouter earns the 10% it adds to the provider's price on paid models, and it earns nothing on your plans. Price only decides between models that already clear the quality bar, and your plans come first. ### How is AutoRouter different from an app's own Auto mode? With AutoRouter you can see which model handled every turn and what it cost, change it with a rule or the app's model picker, and use models from every provider. [Cursor Auto mode vs Copilot auto vs AutoRouter](/blog/auto-mode-vs-autorouter) compares them side by side. ### What does AutoRouter keep to review its picks? AutoRouter keeps the message that opens each turn, up to 4,000 characters, along with records of the model chosen, tokens and cost. Requests pass through AutoRouter's servers on their way to each provider, and AutoRouter doesn't store transcripts. Details are in [AutoRouter's privacy policy](https://autorouter.com/privacy). ### Which apps does AutoRouter work with? AutoRouter works with Codex (the app and the CLI) and the Claude Code CLI on macOS 14 or later, and more apps are on the way. Claude Code needs Node 22.18 or later. AutoRouter is in beta. ## Sources - [AutoRouter](https://autorouter.com) - [AutoRouter privacy policy](https://autorouter.com/privacy) - [Anthropic: Pricing](https://platform.claude.com/docs/en/about-claude/pricing), read October 22, 2026 - [Anthropic: Choosing the right model](https://platform.claude.com/docs/en/about-claude/models/choosing-a-model), read October 22, 2026 - [OpenAI: API pricing](https://developers.openai.com/api/docs/pricing), read October 22, 2026 - [Ong et al., RouteLLM: Learning to Route LLMs with Preference Data](https://arxiv.org/abs/2406.18665), arXiv, June 26, 2024, revised February 23, 2025 ## Read next - [Does model routing make your results worse?](https://autorouter.com/blog/does-model-routing-hurt-quality.md): Model routing makes results worse when it sends work to a model that can't do it. In its October 7, 2026 launch post, Anthropic reported 39.2% for Claude Haiku 5.5 on Terminal-Bench 4.0, a test of long multi-step terminal work, against 70.6% for Claude Sonnet 5.5. AutoRouter, the Mac app for Codex and Claude Code, keeps quality up by making every model clear a quality bar for the task before price counts, sending hard work to the strongest model your plans cover and never treating unclear work as a quick task. - [Which model to use for which coding task](https://autorouter.com/blog/which-model-for-which-task.md): Use a frontier model such as Claude Opus 5.5 or GPT-6 Astra for multi-file refactors, hard bugs and UI built from a design. Use a strong model such as Claude Sonnet 5.5 or GPT-6.1 Sol for tests and code review, and a fast model such as Claude Haiku 5.5 or GPT-6 Luna for renames, commit messages and one-line shell commands. AutoRouter, the Mac app for Codex and Claude Code, makes that choice on every turn. - [What is an LLM router?](https://autorouter.com/blog/what-is-an-llm-router.md): An LLM router is software that sits between an application and several large language models and decides which model handles each request, usually to lower cost or keep requests flowing without giving up quality. Routers range from gateways that pick a provider by price and uptime to rule sets, learned classifiers, cascades and plan-aware routers. AutoRouter, the Mac app for Codex and Claude Code, is a plan-aware router and is unrelated to LiteLLM's Auto Router or OpenRouter's Auto Router. Tags: [AutoRouter](https://autorouter.com/blog/tags/autorouter.md), [Model routing](https://autorouter.com/blog/tags/model-routing.md), [Models](https://autorouter.com/blog/tags/models.md) Drafted with AI from the sources linked above, then checked and edited by the AutoRouter team.