Choosing a model

Start with the default. Change the model when you can name the tradeoff you are trying to improve.

Start with the task

Model choice becomes easier when the task is specific. “Use the smartest model” leaves the requirement vague. Define a concrete outcome, such as extracting decisions from a meeting and turning them into three reviewable tasks.

A task tells you what to evaluate. Does the answer need careful reasoning, fast iteration, a long context window, image understanding, structured output, or a particular writing voice? Once the requirement is clear, model selection becomes a comparison grounded in the work.

Dexto gives you a single place to work across language, image, video, and audio models. You can start with the default model, then switch when the work gives you a concrete reason to do so.

The four tradeoffs that matter

• 01 Capability

  Can the model follow the instructions, reason through the task, and handle its inputs?

• 02 Speed

  How quickly does the task need to move, and how much iteration will the work require?

• 03 Cost

  Is the task exploratory, frequent, or high-value enough to justify a more capable model?

• 04 Modality

  Does the model need to understand or create text, images, video, audio, or files?

These dimensions are connected. A model that is faster may be easier to use for brainstorming. A model with stronger reasoning may be worth the wait when it is planning a multi-step change. A model that understands images may remove a manual transcription step altogether.

There is no universal ranking. There is only a better fit for the work in front of you.

A simple selection loop

Use this small loop when you are unsure:

1. Write the task as a concrete outcome. Include the inputs, constraints, and what a good result should contain.
2. Run it with the default model. The default is a useful baseline because it keeps the rest of the agent setup unchanged.
3. Name the failure mode. Was the result too slow, too shallow, too expensive, poorly formatted, or unable to work with the input?
4. Change one variable. Try a different model or adjust the prompt, skill, or available tool. Change one variable at a time.
5. Keep the simplest option that clears the bar. A A consistently fast model can be the better production choice when it clears the quality bar.

This keeps model selection empirical. Test the system you actually use. Benchmark scores can differ from your work.

A starting map

Use this as a first guess, then validate it with a real example:

The model is only one part of the result. A clear agent, the right skill, useful connections, and a well-scoped project can matter as much as the model itself.

When to bring your own key

Dexto also supports BYOK connections, so you can bring an existing provider subscription or API key into your workflow. This can make sense when you already have a provider relationship, need a particular provider's model, or want to keep usage attached to an existing account.

BYOK still requires a fit evaluation. Compare the result, latency, and workflow behavior on the tasks that matter. It gives you another way to access the model that fits the work.

Model choice is part of agent design

A model picker is one part of the decision. The best agent setup makes model choice work with the rest of the harness:

• an instruction tells the agent what “good” means;
• a skill provides a repeatable method;
• connections provide the information and actions the task needs; and
• approvals keep consequential actions reviewable.

Change the model only after you know which part of the experience needs to improve. Sometimes the answer is a stronger model. Sometimes it is better context, a clearer instruction, or a tool that removes a manual step.