GPT-6 Astra for Ad Creative: What It Runs, What It Costs, and What It Still Cannot Make

TL;DR

OpenAI started rolling out GPT-6 Astra on September 3, 2026. The headline capability for advertisers is computer use: Astra can drive a browser, work through Ads Manager, pull exports, and fill forms across a long task without step-by-step direction. It reads images and writes text, and it does not output images or video itself. So it can run your reporting and your angle development, and it cannot produce the 40 on-brand, correctly sized variants your Advantage+ campaign needs on Monday. Pair it with a creative-production tool like Tadka and you cover both halves.

GPT-6 Astra reached a limited set of organizations first, then ChatGPT Plus, Pro, Business, and Enterprise users, the OpenAI API, Microsoft Azure, and AWS Bedrock. OpenAI put it at state of the art on computer use, browsing, software engineering, cybersecurity, science, and professional work, with 99.9% on ARC-AGI-3 and 100% on ExploitBench. It is also the first OpenAI model rated Critical for cybersecurity capability under the company's Preparedness Framework, which is why early access ran through an application-based program called Daybreak.

For a media buyer, one line from that list does the work: computer use. Everything else is a better version of what GPT-5.6 already did for you.

The specs that affect your workflow

The context window matters for one specific job: dropping a full ad-account export into a single prompt and asking what moved. A 1.05 million token window holds months of daily creative-level data without you pre-summarizing it first.

What a run costs

Astra bills at 10 dollars per million input tokens and 50 dollars per million output tokens at standard rates. Three details change the bill more than the headline rate does:

Run the numbers on a real task. Hand Astra a 200,000 token export of creative-level performance and ask for a written read plus 30 new angles, and you use about 30,000 output tokens. That is 2 dollars of input and 1.50 dollars of output, so 3.50 dollars for the round. Push the same request to a 300,000 token export and the whole request reprices: 6 dollars of input and 2.25 dollars of output, or 8.25 dollars. Splitting one oversized export into two requests under the threshold costs less than sending it in one.

Analysis is cheap at these rates. Creative volume is where the money goes, and no token price changes that.

Computer use is the part that is new

Until this release, a model could tell you what to do in Ads Manager. Astra can go and do it. OpenAI demonstrates it filling out forms, updating CRM records, and organizing calendars, and launch-week demos have it planning and executing multi-step jobs from one open-ended prompt.

The jobs worth pointing it at first:

• Reporting runs. Log into the ad platform, pull the creative-level breakdown for the last 14 days, and write up which hooks held hook rate and which ones decayed.
• QA passes. Check every live ad against your naming convention, tracking parameters, and placement settings, then list what is wrong.
• Competitor sweeps. Work through the Meta Ad Library for six competitors and cluster the angles they are running.
• Angle development. Turn the research above into 30 to 40 distinct messaging angles, each with a hook and a reason it should work for a named audience.
• Post-test synthesis. Read the results of the last test round and write the brief for the next one.

Each of those is a task you currently pay a human to do on a Tuesday morning. None of them puts a new creative in the auction.

Where Astra stops

Astra is a text-output reasoning model with image understanding. It does not natively output images, video, or audio. It can call an image-generation tool through the Responses API, which gets you a picture, and that is a different thing from getting ad creative:

• A picture is not a variant set. Advantage+ and Performance Max want many on-brand assets sized for every placement, refreshed on a cadence. One-off generations do not add up to that.
• Brand consistency is a production problem. Palette, logo placement, type, and safe areas have to hold across every asset. Prompting for them one image at a time drifts.
• The volume math is unchanged. Starving the platform of fresh assets is still the fastest route to creative fatigue, and a smarter agent does not fill the creative supply gap.

The workflow that uses both

1. Point Astra at the account. Have it pull the last 14 days of creative-level results and name the hooks, formats, and offers that held attention.
2. Ask for angles, not copy. Turn that read into 30 to 40 distinct angles with a hook and a one-line rationale each, then rank them by fit to your audience.
3. Produce the variants in a creative tool. Feed the winning angles into Tadka and get a grid of audience-tuned image and video variants that keep your palette, fonts, and logo in place.
4. Ship to the platform. Push them into Meta Advantage+ campaigns or Performance Max asset groups and let the auction sort them.
5. Send Astra back in. Two weeks later it pulls the new numbers, and round two starts from evidence instead of a blank page.

The loop keeps a reasoning agent on the parts that are reading, deciding, and clicking, and keeps a production tool on the part that is making 200 assets that look like your brand.

Which tool for which job

• Use GPT-6 Astra when the task is reading, deciding, or operating: performance analysis, research, angle generation, scripts, QA, and platform admin.
• Use a creative-production tool when the task is output: many on-brand variants, correct sizes for every placement, and refreshes that beat fatigue.
• Use both, in that order, on every campaign where you need strategy and volume, which covers most performance work.

What to do this week

• Give Astra one reporting job you already run by hand and time how long the write-up takes you to check.
• Split any export over 272,000 tokens into two requests before you send it, so you stay on standard rates.
• Ask for 30 angles from your best-performing month, not 3, then cut the list yourself.
• Route production to a tool built for creative volume rather than prompting for images one at a time. Our roundup of the best AI ad creative tools covers the options.
• Track hook rate on whatever ships, and hand those numbers back to Astra for the next round.

Sources: OpenAI: GPT-6 Astra, OpenAI API: GPT-6 Astra model page, CNBC: OpenAI announces rollout of GPT-6 Astra, OpenAI: GPT-6 Astra system card

Astra gives you the read and the angles. Tadka turns those angles into a grid of on-brand, audience-tuned ad creatives across every size your campaigns run, so the platform never runs short on variety. Try it in the studio, or set the loop up with the volume playbook.