Savings calculator

Each task is one agent session of about 324,000 tokens, six turns of about 54,000, mostly cached input, as in the example week on the homepage. Each card prices those sessions at the models it lists, at API list prices. Against Claude Fable, AutoRouter’s hard tasks are priced at Claude Opus, the stronger model for agent work.

How much more work the same spend on your ChatGPT or Claude plans covers. On Free, every task still runs on your plans, just on cheaper models. On Pro, most everyday and quick tasks run on open-weight models paid from Pro’s included usage, so they don’t touch your plans at all.

No. Real routes follow your plans and your rules, and a task goes to a model that does it well before price counts. The models here only set the prices.

Hard: a refactor, a bug across files, a design. Everyday: writing tests, fixing a failing test, a small feature. Quick: a commit message, a rename, a question about the code.

Pro can use the open-weight models AutoRouter offers, including GLM, Kimi, DeepSeek, MiniMax, MiMo and Muse Spark. Here each kind of task is priced at the lowest-cost one rated for it, plus 10%, the way AutoRouter picks among models that clear the quality bar.

Pro is $20 a month and includes $20 of usage. As in the homepage’s example, it sends 75% of everyday and 85% of quick tasks to open-weight models, billed from that usage. The rest stays on your plans, and usage past $20 is added to the month.

No. On a plan you pay the plan’s price, so the saving shows up as the plan lasting longer. See pricing for what each AutoRouter plan includes.

Yes, for five Claude Code tasks run three times on Claude Haiku, Sonnet and Opus 5.5: every run is in results.csv, with the method in method.md.