GPT-6 Astra for Real Estate: Price, Access and What It Can't Do

What GPT-6 Astra Actually Is

GPT-6 Astra is OpenAI's flagship model, launched on 3 September 2026 and rolled out to ChatGPT, the OpenAI API, Microsoft Azure and AWS Bedrock over the following days. Its headline skill is operating software. OpenAI's pitch is "Anything you can do on a computer, Astra can do for you. Fast": clicking through web interfaces, filling in forms and moving between applications on your behalf.

The essentials, checked on 2 October 2026:

OpenAI president Greg Brockman closed the launch briefing with "Welcome to the AGI era." The narrower question for a property business is which parts of your week Astra touches and which it leaves where they were.

Since launch: OpenAI paused new sign-ups to its $200 ChatGPT Pro plan (10 September), added the cheaper GPT-6.1 Sol and GPT-6 Luna, and cancelled GPT-6.1 Astra (28 September). Each change is covered below.

GPT-6 Astra vs Google's Project Astra: The Name Confusion

Two unrelated products share the name, so check which one a vendor or article means.

They are not versions of each other, not a partnership and not competing releases of the same thing. If a vendor pitches "Astra-powered" software, ask which Astra. If it is Google's, ask how they got access to a prototype that Google limits to trusted testers.

How to Use GPT-6 Astra: ChatGPT, API or Cloud

1. ChatGPT, for individual use. Plus includes standard Astra in ChatGPT Work and Codex. Pro, Business and Enterprise plans also get GPT-6 Pro (also called GPT-6 Astra Pro) in regular chat; Enterprise admins have to switch it on. Free and Go plans have no Astra access, though since 22 September they can use the smaller GPT-6 Luna in the desktop app.
2. The OpenAI API, for daily workflows. Call gpt-6-astra and pay per token. This suits agents that run daily, connect to a CRM or call other services such as an image API.
3. Microsoft Azure or AWS Bedrock, for firms already on those clouds. It is the same model, billed through your cloud account. Bedrock's global endpoint matches OpenAI's prices; its in-region and geographic endpoints add 10%, as do OpenAI's own regional-processing endpoints.

What It Costs: API Rates and ChatGPT Message Caps

Two pricing worlds apply: tokens through the API, message allowances through ChatGPT.

API prices

OpenAI's list prices for gpt-6-astra per million tokens, checked 2 October 2026:

Cached input costs a tenth of the input rate. Three things follow:

• Output is the expensive half. A listing description with 8,000 input tokens of context and 900 output tokens costs about $0.13 at standard rates, or about $0.06 through Batch.
• Watch the 272,000-token line. Above it, the whole request is billed at long-context rates, as AWS's model card spells out. A 270,000-token lease pack with a 2,000-token summary costs about $2.80; at 300,000 tokens, about $6.15. Split very large packs first.
• Budget agent runs per session, not per message. Screenshots and tool results pile up in the context, and reasoning effort matters: Artificial Analysis estimates Astra's cost per task at $1.54 at medium effort and $3.26 at max.

Cheaper models in the same family

Scores and cost estimates from Artificial Analysis at max reasoning effort, 2 October 2026:

The listing description above costs about $0.03 on GPT-6.1 Sol. Run routine drafting and extraction on a cheaper model, keep Astra for long multi-step sessions, and test on your own tasks whether the gap is worth paying for.

ChatGPT message allowances

Allowances as reported on 5 September:

On the $100 Pro and Business plans, the GPT-6 Pro allowance is shared with GPT-5.6 Sol Pro. OpenAI says actual usage depends on model choice, context length, reasoning, tool use and caching. If several people will run agent workflows daily, use the API, where cost is metered rather than rationed. On the $200 tier, anyone who cancels or downgrades during the pause cannot re-subscribe until OpenAI lifts it.

What Astra Can Realistically Automate in a Property Business

The real advance is reliability across long, multi-step tasks. In a property business that maps onto an unglamorous list, and every item produces something a person can check.

Listing copy and description drafting

Turning property attributes, floor area, neighborhood notes and a shot list into a description is not new. What changes with a 1,050,000-token context is how much reference fits in one request: your style guide, a few hundred past listings, your compliance wording and your MLS's rules, so drafts sound like your office. Stay under 272,000 input tokens to keep standard rates, and check every fact before publishing; California's regulator says an AI tool does not excuse an inaccurate ad.

CRM data entry and pipeline hygiene

This is where computer use earns its keep. If your CRM has no usable API, an agent can work the web interface directly: open a lead, paste the inquiry, set the stage, schedule the follow-up. Treat it as a clerical assistant with credentials, and keep it to work an unlicensed assistant could do.

Scheduling and viewing coordination

Coordinating a seller, a buyer, an agent and access keys is a bounded, rule-heavy problem with a clear success condition, the shape agentic models handle well. The constraint is permissions: decide in advance what the agent may confirm alone and what it queues for a person.

Inbox triage and first-response drafting

Sorting inquiries by intent, pulling the matching record and drafting a reply for approval is a defensible use today; sending without review is not. In California, a first-contact solicitation must carry your name, license ID number and responsible broker even when AI wrote it, so put those details in the template.

Document reading: leases, surveys, disclosures

Long-context reading is the least exciting win and the easiest to check: summarizing a 90-page lease pack, splitting a bundle into separate documents, flagging non-standard clauses and extracting dates into a table.

Market research and comparables assembly

Pulling comparables, checking listing histories and assembling a pricing rationale is a browsing task with verifiable outputs. Astra's knowledge stops at 30 April 2026, so newer data must come from live browsing or your own records, and every figure needs a source before it reaches a client.

Beyond brokerage

The same pattern fits property managers triaging maintenance requests, hotel and short-term rental teams keeping listings and guest replies consistent, and design or architecture studios turning meeting notes into briefs and specifications. The model handles text and software; visuals are a separate job.

What Astra Cannot Do: The Image Gap

GPT-6 Astra returns text only. AWS's model card lists text and images as inputs and text as the only output. Astra can look at a listing photo and tell you what is wrong with it; it cannot hand back a corrected one. Staging an empty room, brightening a dim kitchen, turning a daytime exterior to twilight, removing a bin from a driveway or redrawing a sketch as a floor plan are image jobs.

OpenAI sells image generation as a separate family: GPT Image 2.5 (gpt-image-2.5-flare and gpt-image-2.5-sunburst) has its own prices on the same list. So if ChatGPT gives you a picture during an Astra session, a different model drew it. What an agent can do is call an image tool for you, as our guide to AI agents and real estate media explains.

What each kind of tool can be asked to do (not how well it does it, so test any tool on your own photos):

In short, Astra fits the coordination layer of a property marketing workflow, not the rendering layer. Whether a general image model can cover rendering is a nuanced question we work through in "Can GPT-6 Astra Do Virtual Staging?".

Where a property image tool fits

Roomagen's tools run on Google's Gemini image models, so the difference from a general model is the workflow around it, not a different model:

• Task questions. Staging asks for the room type and style. Day to Dusk (Time Shift Interior / Time Shift Exterior in the app) has a 24-hour time slider; the interior version asks whether the photo has windows, and its window view defaults to "Keep Original".
• Prompts written for the job. The interior version asks the model to switch on the lights already in the room and keep what is actually visible through the glass.
• Corrections and volume. Refine accepts one written correction (up to 500 characters) on staging, Empty Room, Remove Item and Time Shift Interior results. Paid plans run one tool on 3, 7 or 20 photos per batch, and the Library files each result with its source photo.
• Chat, if you prefer it. The AI Design Chat takes edits in plain words; advice is free and each edit costs 2 credits.

Results usually arrive in under a minute, and paid plans start at $12 a month for 60 credits (pricing). You still check every result and disclose edits where your market requires it.

The Risk Column: Prompt Injection, Cyber Classification, Benchmark Noise

Four things to know before pointing an agent at a live inbox.

Prompt injection is reduced, not solved. Reported figures put Astra's defense against direct prompt injection at 99.99%. Instructions hidden in documents are harder to stop: in Gray Swan's test of 1,810 curated attacks, with 15 attempts per scenario, Astra was compromised at least once in 8.5% of scenarios (GPT-5.6 Sol: 27%). Against jailbreakers who adapt over several rounds, its defense rate falls to about 67%. OpenAI ran those jailbreak tests on the bare model, without the classifiers its products add, but the rule stands: never give an agent both untrusted input and unsupervised write access to money, contracts or published listings.

OpenAI rates Astra "critical" for cyber capability, its first model in that tier: with the right tools and access, it can find unknown vulnerabilities and build exploits without step-by-step guidance. OpenAI also says Astra's written reasoning is harder to monitor than its predecessor's. Give an agent the access controls you would give a new employee, not an admin password.

Its successor was pulled. On 28 September OpenAI cancelled the release of GPT-6.1 Astra; its head of safety systems said the model fell short on "scope and authorization, and how it communicates back to the user about the type of work it's done". In OpenAI's scope test, GPT-6 Astra exceeded its authorized target 0% of the time (GPT-5.6 Sol: 48%), so behavior can change sharply between versions. Re-test whenever the model behind a workflow changes, and check what the agent did in the CRM or outbox rather than trusting its summary.

Benchmarks keep moving. At launch, OpenAI's results had Astra leading in math, coding and cybersecurity. On Artificial Analysis's leaderboard on 2 October 2026, Astra at max effort scores 53, behind Anthropic's Claude Opus 5.5 (58) and Claude Sonnet 5.5 (56), and level with Claude Fable 5.1 and Google's Gemini 4 Argon. Run a five-task trial on your own work before signing anything.

Compliance: Who Is Responsible When the Agent Gets It Wrong

An agent changes who does the work, not who answers for it. California's Department of Real Estate (DRE) said so in a March 2026 advisory:

• Responsibility for inaccurate information, misleading advertising or improper communications from an AI tool "rests with the licensee and their responsible broker, not the technology provider."
• A broker's duty to supervise extends to AI-powered software, and using AI to conduct licensed activity may be equivalent to asking an unlicensed assistant to do it, which violates California law.
• Facts in AI-written ads must be independently verified.
• Since 1 January 2026, digitally altered listing images need a clear disclosure and access to the original (Business and Professions Code section 10140.8).

Disclosure attaches to the edited image, not to whoever triggered the edit, so an edit made through an agent needs the same label as a manual one. Rules differ by state, MLS and country; our AB 723 and MLS disclosure guide covers them in detail.

How to Start Without Betting the Business

1. Pick one task with a checkable output, such as comparables or lease summaries.
2. Run it read-only for two weeks. No sending, publishing or CRM writes; you are measuring accuracy, not saving time yet.
3. Log the failures, including any gap between what the agent says it did and what it actually did.
4. Separate untrusted input from write permissions. An agent that reads incoming email should not hold the credentials that publish a listing.
5. Keep image work on dedicated tools, connected through a real estate image API. Roomagen's API takes one image and one tool per job, can report back by webhook, accepts an Idempotency-Key header so a retry does not create a duplicate job, refunds failed jobs automatically and includes 50 free images on your first key (API quickstart).
6. Budget by session, not by seat. Set a hard monthly spending cap on day one, and try GPT-6.1 Sol before paying Astra rates for routine steps.
7. Write down who approves what before you automate, including which messages carry license details and how edited photos are labeled.

A starting permission map:

The Bottom Line

GPT-6 Astra is a real step in one direction: operating software. For property professionals, that puts the administrative layer within reach of automation (CRM hygiene, scheduling, document reading, first drafts and research), provided a person stays between the agent and anything irreversible. A month after launch, OpenAI's cheaper GPT-6-family models cost a fraction as much for routine steps, and the cancelled GPT-6.1 Astra is a reminder to re-test every model change.

It changes nothing about how a photograph becomes a listing image. Astra writes text; pictures come from image models, and the workflow around them belongs to tools built for that job. The teams that get the most from the Astra era will draw that line clearly.

Prices, plan limits and rankings change often. Figures were checked on 2 October 2026 against the sources below.