Primary navigation

Model selection

Find the right model for your use case

Meet the models

ZeroTwo is multi-vendor. The model picker in the composer lists every model you can choose, grouped by the company that makes it, with each provider's mark beside its models. Try it below: search, jump to a provider, or pick a model to see its details.

Two groups come before the provider sections:

  • ZeroTwo holds Auto, ZeroTwo's own router. Auto picks a model for each message, and your plan includes a monthly allowance of Auto requests.
  • Suggested lists fast, low-cost models that suit most everyday work. On the Free plan this group is replaced by the models Free includes, listed under ZeroTwo beside Auto. Other models use credits on Free.

Each provider section after that lists its models in catalog order. In the picker above, the figure beside each model is its context window, the amount of text it can read at once; in the app that spot can show a credit cost instead. Search matches model and provider names. The sort button beside the search box orders the list by Recommended, Price, Favorited, ZeroTwo preferred, Name or Provider, and each model's row menu has Favorite, Move to top of provider and Disable. See models for plans and limits and pricing for credit costs.

Find the right model for your workflow

Pick a kind of work and a task, then slide toward lower cost or higher quality.

When to consider GPT 6.1 Sol

Consider GPT 6.1 Sol when a task is open-ended and a Luna run keeps falling short, for example turning a financial summary into a board presentation or building a website from a product brief. Run the same input on GPT 6 Luna first and compare, so you can tell whether the extra cost buys anything for that job.

GPT 6.1 Sol is a premium model, so it shows a locked or credit chip on the Free plan. See models for the list of recommended models and the Free roster.

How the recommendation is built

Every task is matched to the capabilities it needs, in the words the catalog uses for each model, such as "coding and programming" or "visual understanding". Only models that list all of them are candidates.

  • Lower latency & costs picks the candidate with the lowest catalog price and stays on the lower rungs, which the catalog describes as fast and economical.
  • Higher quality outputs picks the newest candidate the catalog calls a flagship, or the newest premium one if none does.
  • Balanced also takes the lowest price, except for open-ended tasks such as migrating a service or producing a polished deliverable, where it takes the flagship.

The reasoning effort starts at the model's default rung, rises one rung for open-ended tasks, then moves one rung down or up at the ends of the slider. It only lands on a rung the model really lists, so a model with a short ladder shows fewer choices and one that always reasons shows none. The credit figure is the same per-1k-token estimate the model picker's cost hover gives, worked out from the catalog price. Models your plan includes can spend none, so read it as a relative cost.

This is a rule over catalog listings, not a benchmark. The catalog holds no quality scores, so "higher quality" means the flagship tier, and several tasks share an answer because the catalog lists the same capabilities for them. Use the result as a place to start and check it on a task you know well.

There is no 3D work type: the catalog lists no 3D capability for any model, so there is nothing to base a recommendation on, and Design covers pages and interfaces only. Controlling apps on your own computer, which is what Computer use means here, is a desktop-app feature.

How to think about models and reasoning effort

Start with the cheapest option that could plausibly do the job and move up only when results fall short. In the model switcher, the credit-cost chips make the price of each step visible before you send.

Reasoning effort is a second dial. In Work and Code, GPT 6.1 Sol offers Low, Medium, High, X-High and Max, and GPT 6 Luna offers None through Max. Plain Chat shows only None, Standard and Extended. Low suits small edits and simple extraction; higher rungs suit planning-heavy or conflicting-evidence tasks, at the cost of time and credits.

Experiment

Treat any pairing as a starting point. Run the same input at two settings or on two models, compare the output with what you need, and keep the lightest combination that is good enough.