Best AI model: how to pick one for each task

No AI model is best at everything, because the models at the top of the leaderboards often cost the most and most of a team's work doesn't need them. On October 30, 2026, the Artificial Analysis leaderboard, which aggregates intelligence benchmarks, put Gemini 4 Carbon, Claude Opus 5.5 and GPT-6 Astra at the top. In the Standards model catalog published the same day, Opus 5.5 costs $4 per million input tokens and $20 per million output tokens. OpenAI's Luna 6 costs $0.10 and $0.50, forty times less, and the catalog qualifies it at standard level on all five tasks. Summarizing a client meeting or copying a policy number out of a PDF doesn't call for the top of the class. That leaves three things to settle: the task, how hard it is, and the cheapest model that can do it.

Every request you give an AI falls under one or more of five tasks, and the task alone rules out part of the field. The Standards catalog defines them this way, and its Auto mode sorts each message into these five before it picks a model.

One request can mix two tasks: pull the figures from a bank statement, then write the summary for the client. The model you pick then has to be qualified for both.

Demand measures how much reasoning a task takes, and the Standards catalog uses three levels. A light task is one short step with no ambiguity. A standard task takes several steps or some judgment. A heavy task is long, ambiguous, or needs sustained reasoning. The catalog qualifies each model, task by task, up to a level, and that level works as a ceiling: a model qualified heavy for writing also handles standard and light writing, but not the other way round. In practice, most of a firm's requests in a day are light or standard: find a date, summarize a call, fill in a record. Heavy tasks exist, like an estate plan with several scenarios or a file full of contradictions, but they're rare, and they're the only ones that need a model qualified heavy. Heavy doesn't have to mean expensive: on October 30, 2026, the catalog qualifies heavy models at every price point.

Once you know the task and its level, take the cheapest model qualified at that level, and don't reach for the strongest one by habit. The table below applies the rule to a common job in a firm, writing at standard level. It's read live from the Standards catalog and sorted from cheapest to most expensive.

Standards' Auto mode applies this rule. The app describes it in one line: "Picks the model that fits each message: chat, writing, code, analysis or extraction, and how demanding it is." Among the models qualified for every task in the message at the level it needs, Auto takes the cheapest. It does so under a price ceiling of $5 per million input tokens and $25 per million output tokens when a person is following the conversation. It never drops below the level to save money, and when the conversation carries on the same work, it keeps the model already qualified. You turn Auto on from the model menu, or you can set a fixed model for an agent or a conversation, from OpenAI, Anthropic, Google, Mistral or the other providers in the catalog.

For code, take the cheapest model qualified heavy if the task is long or ambiguous, and standard otherwise. On October 30, 2026, nine models in the Standards catalog hold heavy level for code. Here are the five cheapest, cheapest first.

In an advisory firm, code stays occasional: a spreadsheet formula, a script that imports a file of policies. A simple formula is a light task; an import script that has to handle messy edge cases is heavy.

The last step is to look at what your tasks cost, since a price per million tokens doesn't tell you how many tokens your requests use. Standards shows the estimated cost of every answer and tracks AI usage by day, by model and by team member. After a week or two you can see which requests go to an expensive model and whether they deserve it. An agent that writes a morning brief from dozens of records reads a lot of tokens. Moving it to the cheapest model qualified at standard level cuts its bill, if the task was standard all along. If an analysis keeps coming back with reasoning errors, that tells you the opposite: the task was heavy. Change the model for that one agent.

A fictional twelve-person wealth management firm that sends four kinds of request to an AI shows the rule at work: light conversation, standard writing, standard extraction and heavy analysis don't need the same model. The models named are the ones in the Standards catalog on October 30, 2026; they'll change.

The first three requests make up most of the firm's volume, and they run on light or standard models, from $0.10 per million input tokens. Only the fourth needs a model qualified heavy for analysis. On October 30, 2026, these are the models the Standards catalog qualifies heavy for analysis:

On October 30, 2026, the Standards catalog lists 23 text models from eleven providers. Anthropic, OpenAI, Google and Mistral account for ten of them, and the table gives, for each provider, what the catalog says, as applicable: default model, recommended models, levels qualified.

DeepSeek, Z.ai, Alibaba, Moonshot, MiniMax, Meta and SpaceX AI complete the list. GPT-6 Astra and Gemini 4 Carbon, top of the performance leaderboards, aren't in it. These profiles move: by default, every Standards deployment rereads the catalog every 30 minutes, and a model added there becomes available with no product release.

For what an agent may do on its own once its model is set, read our article on AI agent security for firms that hold client data. For the API and the SDK, see the developer page.