← All Perspectives

OpenAI released GPT-6.2 Astra and Sol. Which would you give the job?

Both models can work with the same customer notes. A quick summary and a presentation for a client give you different reasons to choose one.

Interview cards sit behind a short summary on the left. Reports and a spreadsheet stand behind a slide deck on the right.
The same customer interviews can lead to a quick summary or a presentation that makes a case for changing the product.

On November 17, OpenAI released two GPT-6.2 models. The way I’d choose between them starts with a question you might hear twice in the same week. On Monday someone hands you a stack of customer interviews and asks, “What keeps coming up?” On Friday the client asks, “Can you show us why we should change the product?” You might reach for the very same notes. I’d have a different model in mind the second time. OpenAI’s release announcement (opens in a new tab)

From a stack of notes to a client presentation

For Monday’s question, I’d start with Sol. Ask it to group the recurring complaints and pull a quote that illustrates each one. You know what a useful answer looks like, and you can read it beside the notes. That’s the kind of everyday, multistep work OpenAI designed Sol to do quickly.

Friday’s request gives Astra a reason to earn its keep. Say the client has a slide template, the interviews tell you where people get stuck, and a spreadsheet shows how often it happens. Someone has to bring those sources together and explain what they support. OpenAI reports that Astra follows supplied templates more closely than its predecessor on work like this. That makes the deck a job I’d try with Astra. OpenAI’s description and evaluations (opens in a new tab)

The slide I’d look at first is the one the client will use to make a decision. If it says a problem is getting worse, can you trace that line back to the spreadsheet? There’s a reason to ask. In two of OpenAI’s evaluations, the new models were more likely to leave out obstacles they’d run into, including a missing file. OpenAI says a change to its Codex instructions helped. In our deck example, I’d want the model to tell me if it never got the spreadsheet. A client could act on that slide. OpenAI’s reliability findings (opens in a new tab)

A line connects a chart in a source report to the matching chart on a presentation slide.
A reader should be able to trace the claim on a slide back to its source.

What the price difference buys

If your software calls the API, you pay for the text you send and the text you get back, measured in pieces called tokens. Suppose one request uses 10,000 uncached input tokens and produces 2,000 output tokens. At standard short-context rates, that’s 4 cents with Sol and 16 cents with Astra, before any tool fees. That’s a price example. I haven’t run the interviews or the deck through either model. OpenAI API pricing (opens in a new tab)

Twelve cents is an easy difference to see. Editing time is the part I’d want to learn from you. If Sol gives you a sound summary on the first pass, you can put it to use. If Astra hands you a deck that already fits the client’s template, the time you would have spent rebuilding slides belongs in the comparison too.

If you use ChatGPT Work or Codex through your ChatGPT plan, those per-request prices apply to a different way of buying access. Your work draws on the plan’s shared allowance and, where applicable, credits. The models you see depend on your plan, seat, workspace settings and rollout. Look in Work or Codex for this release. Regular Chat keeps its existing picker. Work and Codex guide (opens in a new tab)

I’m curious about the first job where you’d choose the same model again. If you’ve tried Astra and Sol on the same assignment, what did you give them, and what did you have to change before you could use each answer?

Share this perspective

Loading the conversation…

Keep exploring

A useful next step