OpenAI announced GPT-6 Astra on September 3, 2026, and the model id that returned "model not found" a day earlier is now live in the API documentation with pricing, rate limits, and a 1,050,000-token context window. Day-one access is limited — enterprises in OpenAI's Trusted Access Program can call the API immediately, while Plus, Pro, Business, and Enterprise users of ChatGPT get access "in the coming days" — but the capability list is what makes this release different for SEO teams. Astra operates a computer through the same inputs a human uses, and it generates and hosts complete webpages and documents. Both abilities point at the same conclusion: the agentic SEO workflows teams have been prototyping are about to get a model built for them. Status is current as of September 4, 2026, and everything below is from OpenAI's announcement and API documentation unless attributed otherwise.
What OpenAI announced
Item | Detail |
|---|---|
Model | GPT-6 Astra, the successor to GPT-5.6 Sol; OpenAI positions it as "our most capable model, built for the hardest end-to-end work" |
Date | Announced September 3, 2026; staged rollout begins the same day |
Access | API and Codex for Trusted Access Program enterprises now; Plus, Pro, Business, and Enterprise access "in the coming days"; not planned for free or the cheapest paid tier, and Enterprise administrators must enable it |
Context window | 1,050,000 tokens (128,000 max output) |
API pricing | $10 per million input tokens, $50 per million output tokens; cached input at $1. Requests above 272,000 input tokens bill at 2x input and 1.5x output rates for the whole request |
Inputs and outputs | Text and image input; text output; tools include computer use, image generation, code interpreter, file and web search, hosted shell, skills, and MCP |
Knowledge cutoff | April 30, 2026 |
Safety stance | First OpenAI model rated "Critical" in its Preparedness Framework; layered restrictions and a real-time monitor that can slow, pause, or stop agent tasks |
The announcement leans on two numbers that frame the release: OpenAI describes Astra as its largest training run "by far," and it says the model completes OSWorld 2.0 computer-use tasks with 72.6 percent success while cutting average task time from roughly 75 minutes to about 40 — a 47 percent speedup. These are OpenAI-reported figures from the model card and demos, not independent measurements.
The two upgrades that matter for SEO work
Computer use is now pixel-level. Astra operates a screen the way a person does — reading pixels, moving the mouse, typing — rather than calling custom connectors for each site. OpenAI's demos included filling a tax form from a scanned document, updating CRM records, and running front-end QA checks on websites. For SEO teams, the significance is that browser-native tasks no longer depend on a vendor building a connector first: any tool that runs in a browser, from Google Search Console to a competitor's signup flow, becomes a candidate for automation. OpenAI notes computer-use tool calls carry a fee per call; the pricing page for it is not published yet.
Creation is now page-and-document grade. Astra can create, host, and share websites and web apps from a prompt — OpenAI calls the ChatGPT surface "Sites" — and can produce documents and presentations that match a user's existing style. The API adds an image generation tool alongside standard text output. That closes the production loop for content teams: brief, draft, image, and a hosted page can come out of one agent run, which is precisely the pipeline the SEO industry has been approximating with disconnected tools.
What this changes for SEO and GEO teams
1. Link work moves into real browsers. Outreach forms, directory updates, profile creation, and index submission all happen in logged-in web interfaces that have historically resisted API automation. A pixel-driven agent that can read a captcha-free form, fill it, and verify the result turns "submit this URL" from a manual task into a scriptable one — with the caveat that every target site still sets its own terms, and automation policies have not changed just because the model got better.
2. Competitor research becomes hands-on. Monitoring what competitors ship no longer has to stop at reading their changelogs. The same agent can sign up for a trial, click through a new feature, and report what the page actually does — then feed that observation into your content and positioning work. The workflow our team has been pointing at for months — dated snapshots of competitor pages and AI answers, diffed on a schedule — is now executable end to end by one agent with browser access and a file system.
3. Production quality raises the floor for generated pages. Mass-generated SEO pages have been under scrutiny all year, and rightly so — much of it is template junk that retrieval systems increasingly discount. Astra's site and document generation raises what an automated pipeline can produce: real layouts, working navigation, original images, and content that looks maintained. The risk is not better pages, it is better junk at scale; the differentiator stays what it always was, and the evidence is checkable in seconds. A page either answers the query with named, verifiable sources or it does not.
4. Audits can act instead of just reporting. OpenAI's own demos show Astra running front-end QA against live websites. For technical SEO that means the gap between "the audit says the schema is wrong" and "the schema is fixed" can close inside one agent session: diagnose a page, edit the template, re-render, and re-check against the original finding. This is the loop our Codex SEO skill series packages — and it is the reason those skills were written model-agnostic. The same skill files that ran on earlier models should run on Astra with the model setting changed and a dated re-baseline.
Where the limits are, as of September 4
The honest list is longer than the hype list. Most SEO teams cannot call the model yet — consumer and Business plan access lands "in the coming days," and Enterprise requires an administrator to opt in. OpenAI's own safeguards can pause long agent tasks mid-run for review, which matters for overnight automation runs. The refusal behavior described in the announcement — the model declines certain advanced cybersecurity tasks in widely deployed versions — has not been fully mapped onto everyday use cases like large-scale crawling. Computer-use tool pricing is unpublished, and cost per completed task is the metric OpenAI itself tells customers to watch, not cost per token. And the company has said monitoring gets harder as models get more capable; it has committed to withholding further scaling if it loses confidence in alignment monitoring. None of that changes what the model can do in a browser today. It changes how carefully a team should checkpoint the jobs it hands over.
The Auspia read
For the teams we work with, this release compresses the distance between SEO strategy and SEO execution. The model that can fill out a form, run a front-end QA pass, and ship a hosted page is the model that can own a weekly SEO operations loop end to end. The discipline that made that loop safe before Astra is the discipline that keeps it safe now: package the job as a skill, baseline it on a small scope, spot-check findings by hand, and re-run the baseline the day the model reaches your account — which, as of this week, is finally a date your team can put on the calendar. We wrote the setup guide before the release; the re-baseline step in it is now runnable.
Author: Jasper Quinn, AI Search Product Researcher Tracking 60+ Feature Changes at Auspia. Jasper tracks model and platform releases and translates what they change for search and content operations.
