GPT-6 Astra for Google and Meta Ads: What It Changes and How to Connect It
GPT-6 Astra shipped to paid ChatGPT users on September 4, 2026. What is confirmed, what to expect for ad-account work, and the ten-minute setup that connects it to a live Google Ads and Meta Ads account.
Built by our community of 2,000 marketers
Free prompts for paid ads
Templates for ChatGPT, Claude and Perplexity.
Clients we work withClients we
work with
Quick answer
GPT-6 Astra is OpenAI’s new model, in ChatGPT Plus, Pro, Business and Enterprise since September 4, 2026, and on the API at $10 per million input tokens and $50 per million output. For advertising work, the change that matters is not writing better headlines; every model since GPT-4 has written acceptable headlines. The change is that OpenAI built this one for multi-step tasks with tool calls, which is the shape of real ad-account work: pull the last 14 days, find the ad groups whose CPA moved, check search terms, propose negatives, apply them, log it.
Disclosure: Ryze AI publishes this page and sells the connector and the autonomous plan described at the end. The connector is free to attach; the $89/month plan is for the part where Ryze runs the account without you. And a con up front: an MCP connection gives GPT-6 the same access a junior buyer would have, so the guardrails section is not optional.
What is confirmed (as of September 8, 2026)
Everything in this section comes from OpenAI’s release material and the coverage of the launch. Nothing here is our own testing.
Release
Limited preview on September 3, 2026; paid ChatGPT users (Plus, Pro, Business, Enterprise) on September 4, 2026; also available through the OpenAI API and Amazon Web Services.
Positioning
OpenAI describes it as its most capable and most aligned model, with state-of-the-art results in coding, math, computer use and browser navigation. The launch notes call out staying focused, adhering to task boundaries, understanding user intent, handling tedious tasks and completing multi-step workflows.
Reasoning
The model uses a recurrent-depth approach, sometimes described as looped transformers. One consequence reported at launch is that the reasoning trace is less legible than in earlier models. For ad work this matters: if you cannot read why it wanted to pause a campaign, you need the change log to carry the reason instead.
Restrictions
The public build declines some cybersecurity prompts; a fuller build is gated to vetted defenders. This does not touch advertising work.
Pricing
API $10 per million input tokens, $50 per million output. A daily ad-account check that pulls a few thousand rows costs cents; a full account audit with search terms runs to a dollar or two. Verify against openai.com/pricing before quoting, because launch prices move.
What to expect for Google Ads and Meta Ads work
This is the expected part, stated as expectations. We will replace it with measured results once our own runs are done; the test protocol is below so you can run it before we publish ours.
Longer procedures complete
On earlier models a ten-check CPA diagnostic tended to skip the boring middle checks on a large export. A model tuned for task boundaries should run the list in order. Expect the biggest gain in scheduled, repetitive jobs (daily anomaly check, weekly fatigue scan, monthly wasted-spend sweep), not in one-off creative brainstorming.
Two-account runs stay separated
Cross-platform prompts pull Google Ads and Meta through two MCP tools in one conversation. The failure mode on older models was mixing metrics across platforms or estimating a missing slice. Expect fewer of those, and keep the line “if a slice is missing, ask, do not estimate” in every prompt anyway.
Computer-use gains mostly do not matter here
Ads Manager and the Google Ads UI are the slow path; the API through MCP is the fast one. Use computer use only for the things the APIs do not expose, such as some Meta creative previews and Google Ads Editor bulk sheets.
Costs drop per task, not per token
Output tokens are $50 per million, so a verbose model is expensive. Ask for tables and decisions, not narratives.
Do not expect
Knowledge of your account without a connection, correct knowledge of 2026 platform changes (AI Max defaults, the DSA sunset timeline) without you pasting the source, or safe autonomous writes on day one.
The same expectations applied to the last Claude release; the measured results for that model are in Claude Fable 5.1 for Google and Meta Ads, and the model-neutral prompt sets are in ChatGPT prompts for Google Ads and ChatGPT for Facebook Ads.
Free SEO/GEO audit
- Surface technical SEO issues
- Spot quick wins
- Find duplicate and thin pages
Free · no credit card · instant
How to connect GPT-6 Astra to Google Ads and Meta Ads
ChatGPT connects to outside systems through MCP apps in Developer mode. The steps below use Ryze AI’s connector, which exposes Google Ads, Meta Ads, TikTok, LinkedIn and ChatGPT Ads tools on one endpoint; a self-hosted MCP server on the Google Ads and Meta Marketing APIs works the same way at step 4. The connector reference is on the ChatGPT ads MCP page.
Turn on Developer mode
In ChatGPT, open Settings, then Apps, then Advanced settings, and turn on Developer mode. This needs a Plus, Pro or Business plan.
Create the app
Choose Create app and name it “Ryze AI”.
Enter the server URL
Server URL: https://connector.get-ryze.ai/mcp. Pick Server URL, not Tunnel. Authentication: OAuth.
Sign in and approve
Sign in with your Ryze AI credentials, choose the workspace, approve. The workspace is where your Google Ads and Meta accounts are already connected; if you have not connected them yet, do that first at get-ryze.ai.
Verify the connection
Start a new chat with GPT-6 Astra selected and ask: “List the ad accounts you can see and the date range of data available for each.” If it returns your account names and IDs, the connection is live.
What the connector exposes
Reads for campaigns, ad groups, ads and conversion events with pagination; insights (impressions, clicks, spend, CTR, CPC, CPM) at each level; writes to create, update, activate and pause campaigns and ad structures, gated behind approval; archive, which is irreversible and requires explicit approval; creative upload from an image URL. Attaching the connector is free. The $89/month plan is for Ryze’s own autonomous optimization, not for the connection.
Con. This is the same connector on every model, so nothing here is specific to GPT-6. The model changes how well the tools get used, not which tools exist.
Five prompts to run in week one
Run each one read-only. Paste the result into a doc so you have a baseline for the next model.
01 · read-only
Overnight anomaly check
Yesterday against the trailing 7-day average, per campaign, across both platforms.
02 · read-only
Wasted-spend sweep (Google)
Search terms with spend and no conversions, grouped by intent, with proposed negatives.
03 · read-only
Creative fatigue (Meta)
Frequency, CTR trend and CPA trend per active ad, with pause or refresh calls.
04 · read-only
Budget reallocation scenario
A modelled move of budget from the worst-CPA campaigns to the best, assumptions stated.
05 · read-only
Client-ready summary
A 150-word weekly summary a non-technical client can read.
Prompts 2 to 4 are the ones where a task-boundary improvement should show. If GPT-6 stays inside “do not apply anything” every time across a week, it has earned the next section.
Letting it act: guardrails that do not depend on the model
Start with one narrow write
Adding negative keywords. It is reversible and low-cost.
Cap the workspace
Put a spend cap on the workspace and a pre-approved action list on the connector. Anything outside the list stays a proposal.
Keep the change log on
GPT-6's reasoning trace is harder to read than GPT-5.x's, so the log with reasons is your audit trail.
Two weeks read-only first
Schedule read-only checks for two weeks before scheduling any write.
No archive rights
Never give an autonomous loop archive rights. Pausing is enough.
Ryze AI’s Paid Ads Autopilot does this loop itself at $89/month, with the approval queue and log built in, and it does not depend on which model you talk to it from. Con: that is our product, it earns autonomy over a baseline period rather than instantly, and it gives you fewer per-check switches than a hand-built script.
The test we will publish (and how to run yours first)
Same account, same 30-day window, the five prompts above, three runs each on GPT-6 Astra, GPT-5.6 and Claude Fable 5.1 through the same connector. Score: did it finish every step, did it keep platforms separate, did it stay read-only when told, how many output tokens did it use, and how long did each run take. We will put the numbers here with the date. Until then, every claim in “what to expect” is an expectation.
The comparison of the predecessors is in Claude vs ChatGPT for Google Ads; the broader picture of ads inside ChatGPT itself is in ChatGPT Ads.
If you want the documentation behind this, the Ryze help center has pages on what the Ryze agent manages in Meta ads and approvals.
Frequently asked questions
Can GPT-6 Astra manage Google Ads?
Does GPT-6 work with Meta Ads Manager?
Is GPT-6 better than Claude Fable 5.1 for ads?
What does it cost to run GPT-6 on an ad account?
Show 1 more questionsShow fewer questions
Will GPT-6 replace an ads manager?
Closing
Treat GPT-6 Astra like a new hire with a good memory and no context: connect it, give it the five read-only jobs, read its work for two weeks, then hand over one reversible action. If you would rather not run that process yourself, Ryze AI runs it for $89/month, with the honest caveats that it is ours and that it, too, has to earn the keys.
