All posts
AI modelsMarketing reportingPaid media

GPT-6.2 Astra vs. Sol: Which should write your marketing report?

Test GPT-6.2 Astra and Sol on the same campaign export. Use a ready-to-copy prompt and compare accuracy, missing data, and how much work each draft takes to review.

Magister Marketing5 min read

Try GPT-6.2 Sol first for a routine weekly marketing report. Then give GPT-6.2 Astra the same export and instructions. Keep the draft that gets the numbers right and takes less work to review. If neither can explain where a number came from, neither is ready to share.

This test is for founders and marketers who already pull campaign numbers and want help writing the update. Use a past report you can check against the original data. The aim is to pick a model for your report, not to declare a winner for everyone.

What changed with GPT-6.2 Astra and Sol?

OpenAI introduced GPT-6.2 Astra and GPT-6.2 Sol on November 17, 2026. OpenAI describes Sol as its model for everyday writing and work that takes several steps, and Astra as its choice for harder projects. In the launch notes, one of OpenAI’s Astra examples is turning a report into a slide deck using your team’s template. That makes Astra worth testing if fitting a strict format is part of the job. It doesn’t tell you which model will calculate your cost per acquisition correctly.

As of November 19, 2026, OpenAI documented both models in ChatGPT Work on supported plans, in the Codex app and CLI, and in the API. In ChatGPT, open Work rather than the standard Chat mode. Enterprise and Edu admins need to enable the models before their teams can select them. Check OpenAI’s current plan and access details if you don’t see them.

Start with a week you can check

Export two comparable reporting periods from one ad account. Use campaign-level spend and the conversion event your team actually reports on. Include the exact dates, account time zone, currency, and conversion definition. If the export includes a total row and individual campaigns, mark the total row so it won’t be counted twice. Use an approved workspace and an aggregate file with no customer records or identifiers. You don’t need to connect a live account for this test.

A calendar week can be over while conversions are still coming in. As Google Ads Help explains (checked November 19, 2026), recent spend may already be reported while some conversions haven’t arrived, making cost per conversion look higher. Pick periods with comparable conversion coverage. If that isn’t possible, label the newer period provisional rather than asking either model to call it a decline.

Before you run either model, make an answer key in your spreadsheet. Check total spend, total conversions, and whether each went up or down. For account-level cost per acquisition (CPA), divide total spend by total conversions for the same campaigns and dates. Don’t average individual campaign CPAs. If conversions are missing or the denominator is zero, don’t report a CPA. The UK Government Data Quality Framework recommends communicating what data covers and how quality issues affect its use. Put those caveats in the report, not just in your notes.

Give both models the same assignment

In ChatGPT Work, start a separate session for each model and attach the same approved export and report template. Use the same instructions. Match reasoning effort and tool access where possible. If those settings differ, record them and treat this as a comparison of the setups you’d actually use, not a test of the models in isolation. Don’t share your answer key or the other model’s draft.

Try this prompt with both models. Fill in the bracketed details from your export and reporting notes.

Draft an internal weekly paid media update for a founder.
Use only the attached export and report template.

Reporting period: [exact dates]
Comparison period: [exact dates]
Account time zone and currency: [details]
Conversion definition: [event counted]
Known gaps or provisional data: [details, or none known]

Start with a short summary. Then show spend, conversions and CPA
for both periods, with the change and where each figure came from
in the file. Calculate CPA from matching spend and conversions.
Do not count subtotal or total rows again when summing campaigns.
Mark missing data as missing, not zero. Where a denominator is
zero or an input is missing, mark the rate as not calculable.

Separate what changed from possible explanations. Flag comparisons
that the file cannot support. End with one specific follow-up check.
Do not change any campaigns or send the report.

Check the numbers before you judge the writing

Start with the answer key. Do the headline totals match? Did the draft compare the right dates? Is the direction of each change correct? Can you trace any campaign singled out in the summary back to the export? NIST’s Generative AI Profile recommends verifying sources and citations in generated outputs. Here, that means opening the file and checking the figures, not trusting a plausible-looking reference.

Watch how each draft handles uncertainty. If a signup column is blank for one period, “signups fell” is not a finding. If CPA did rise, the export might establish the change without showing why. “Check whether late conversions are still coming in” is a useful next step. “Conversion lag caused the increase” isn’t supported by a change in CPA alone.

Drop any draft with a material error. For drafts that pass, track the edits you actually make and the time it takes to get each one ready to share. Note whether the draft used your template and whether its follow-up gives someone a clear thing to check. A polished opening paragraph doesn’t make up for a table you have to rebuild.

When is Astra worth using?

As of November 19, 2026, OpenAI’s Standard API rates for requests with up to 272,000 input tokens were $2 for uncached input and $10 for output per million tokens for Sol, and $8 for uncached input and $40 for output per million tokens for Astra. Those aren’t prices per report. Other processing tiers and longer inputs can have different rates, and API billing is separate from your ChatGPT subscription’s included usage.

If both drafts are accurate and Sol takes about the same effort to review, keep Sol for that report. Try Astra on a report with conflicting inputs or a template that usually needs more cleanup. Keep it for that job if the handoff is genuinely easier and worth the added cost or usage. If both drafts get the numbers wrong, fix the export or clarify the instructions before switching models again.

Run the test on one past report this week. Save the export, prompt, and answer key so you can repeat it when your reporting needs change. The same principle applies to tracking AI visibility: keep the inputs comparable and record what’s missing before you interpret a change.

Drafted by Magister, reviewed by our team.