GPT-6.2 Astra or Sol? How I Pick Models for My Personal AI Agent

How I pick Luna, Terra, Sol and Astra for different jobs in my personal AI agent.

Watch on YouTube ↗Explore the gear →

OpenAI released GPT-6.2 Sol and Astra today. I went through the work my personal agent Joanna already does and moved the existing Sol tasks to 6.2 Sol. I still give email, chat, and the weekly pass through our notes different models. The choice is easier when I look at one task at a time.

WHICH MODEL? beside a phone with a chat icon

Where I use each model

GPT-6 Luna helps flag spam through OpenAI’s Decisions API. It gives my software a classification to work with. That check happens often, so the cost of each pass matters. I also give Luna background work such as looking at a network alert.

GPT-6.1 Terra gets the conversation. If I ask Joanna about a reminder and follow with “when is it due?”, I want her to follow the reference and reply quickly. I also use Terra to pull a date and time out of an appointment screenshot.

GPT-6.2 Sol gets jobs that need more reading and writing. In the weekly memory pass, Joanna can pull together useful notes and keep track of corrections. That’s the kind of work I want her to spend more effort on because I use the result again later.

When Astra would be worth trying

If I need to make sense of several logs and a timeline after something broke, I’d try GPT-6.2 Astra. At standard Codex credit rates, a million output tokens cost 12.5 credits on Luna, 125 on Terra, 250 on Sol, and 1,000 on Astra. The Decisions API is priced separately.

Six point two Sol has the same listed token rates as six point one Sol. OpenAI reports better output-token efficiency. I still want to see how many tokens it takes to get a useful result from my own tasks.

Try it with one recurring task

  1. Choose a task your agent runs regularly and save a handful of representative inputs you can reuse.
  2. Run the same inputs through the models you’re considering. Check whether each result is right, how long it took, and how many tokens it used, including retries.
  3. Pick the model that reliably finishes that task within your time and usage budget. Revisit the choice when the task changes.

If you’ve got your own agent running, I’d love to hear which job made you change models. For another project I built with Codex, here’s what I did with my old Nest thermostat.

Watch Is GPT-6.2 Astra Right for Your AI Agent? on YouTube

Advertisement

Keep building

The next useful idea, in your inbox.

Get new vCloudInfo posts by email.

Get new posts