AGI, GPT-6 Astra and Where AI Is Actually Going

What is AGI, and why did GPT-6 Astra reopen the question?

Artificial general intelligence (AGI) is an AI system that can do most of the intellectual work a capable adult can do, across most fields, without being rebuilt for each task. That is the everyday meaning. The trouble is that the companies closest to building it each define it differently, and the definition decides whether the claim "AGI is near" is exciting, alarming or just accounting.

The question stopped being abstract on 3 September 2026, when OpenAI announced GPT-6 Astra and rolled it out to paid ChatGPT users the next day. OpenAI called it its most capable model yet, and co-founder Greg Brockman said publicly that it could mark the arrival of AGI. OpenAI's own charter defines AGI as a system that can perform all economically valuable work as well as or better than humans, so measured against that yardstick the claim is enormous, and worth reading carefully rather than taking at face value.

Why this matters to a business owner: the useful question is never "has AGI arrived". The useful question is "which specific tasks in my company can a machine now do at a competent-employee level, and what happens to my data and my risk when I let it". That question has a practical answer today, and Astra widened it again.

What is GPT-6 Astra?

GPT-6 Astra is OpenAI's frontier model, released in early September 2026 to ChatGPT Plus, Pro, Business and Enterprise users and through the OpenAI API and AWS, replacing the GPT-5.6 generation. OpenAI describes it as a generational leap in software engineering, science, professional work and, notably, cybersecurity, and reports that it saturates reasoning and mathematics benchmarks that earlier models could barely touch. It handles very long tasks in a single pass, with a context window measured in the millions of tokens.

Two things about Astra matter more to a business than the benchmark scores. The first is that OpenAI says the model meets the "Critical" cybersecurity threshold in its own preparedness framework, meaning that with the right access it can find previously unknown security flaws and build working exploits with little human guidance. OpenAI shipped a restricted version that refuses some of that work, but the capability is now real, and it cuts both ways: defenders and attackers reach for the same tool. The second is that Astra uses a "recurrent depth" style of reasoning that hides much of its working, which is why safety researchers have questioned how easily its decisions can be checked.

A note on the name, because it causes genuine confusion. Google DeepMind runs a separate, older project also called Project Astra, a universal-assistant research effort shown at Google I/O in May 2024 and since folded into Gemini Live. It is a different thing from OpenAI's GPT-6 Astra. When people talk about "Astra" as the new AGI-class model in September 2026, they mean OpenAI's.

How close is AGI in 2026?

Nobody can say with confidence, and anyone who gives you a firm date is selling something, so the honest answer is a range of informed views with names attached. OpenAI now frames Astra as a candidate for AGI, a claim that has drawn as much scrutiny as celebration. Dario Amodei of Anthropic argued in late 2024 that "powerful AI" could arrive around 2026. Demis Hassabis of Google DeepMind has repeatedly put AGI five to ten years out and points to models still lacking consistency, real reasoning and the ability to learn continuously. A single model launch, however striking its benchmarks, does not settle a debate the field itself has not settled.

What is beyond dispute is the trend under those opinions. Each generation of frontier models handles longer tasks with less supervision: in 2023 an assistant could draft an email, in 2024 it could write and debug a working program, and by 2025 agents could complete multi-step jobs in a browser or a code repository over hours, with a human checking the result rather than each step. Astra pushes that line further again. Independent measurements of how long a task a model can finish unaided have roughly doubled every several months, and if that continues the systems of 2027 will handle work that takes a person a full day.

So the realistic planning assumption for a South African business is not "AGI arrives on a date". It is "each year, a wider band of white-collar tasks becomes something software can do competently, cheaply and around the clock". Plan for the band widening, and the exact moment someone declares AGI becomes a headline rather than a shock.

What changes when assistants become agents?

The change is from advice to action: an assistant answers a question, an agent is given a goal, an account and permission, and then does the job. That single shift rewrites most of the risk model a business already has for AI.

• Data exposure becomes structural. A chatbot only sees what someone pastes into it. An agent connected to your mailbox, file shares and CRM sees everything those permissions reach. The question stops being "what did staff paste" and becomes "what can the agent read", which is a question about identity and access control, not AI.
• Mistakes execute. A wrong answer in a chat is caught when a person reads it. An agent that misreads an instruction can send the email, change the invoice or delete the folder before anyone looks. The controls that matter are the same ones you would put on a new junior employee: limited permissions, approval steps for anything irreversible, and logs.
• Perception adds a new surface. When an assistant sees through a camera and hears the room, whatever is on the whiteboard, the screen behind you or the client's paperwork on the desk is in scope. POPIA does not care whether a person or a model processed the personal information.
• Vendor terms decide ownership. Business tiers of ChatGPT Enterprise, Microsoft 365 Copilot, Gemini for Workspace and Claude for Work commit in writing not to train on your data. Consumer tiers generally do not. As agents gain access to more, that line in the contract is worth more than any feature.

What does this mean for a South African SME?

For a company of ten to three hundred people, the arrival of agentic AI is mostly good news with three local complications. The good news is that the productivity gains that used to need an enterprise budget now come bundled into Microsoft 365 and Google Workspace licences priced in rand per user per month.

The complications are these. First, POPIA puts the responsibility for personal information squarely on the business, whatever tool touched it, so an agent reading customer records needs the same lawful basis and safeguards as a staff member would. Second, data residency is now a real choice: Microsoft, Amazon and Google all run cloud regions in Johannesburg or Cape Town, and for regulated sectors it is worth choosing services that keep processing in-country where the option exists. Third, exchange-rate exposure: most AI tiers are dollar-priced, so a per-user add-on that looks cheap in a US blog post costs a South African company meaningfully more, and should be justified by a measured saving rather than enthusiasm.

There is also a security point that Astra sharpens. A model that can find and exploit software flaws with little human help lowers the cost of an attack as much as the cost of a defence, so the ordinary basics stop being optional: multi-factor authentication, prompt patching and tested backups are what stand between a small business and an automated attacker. That is the same cybersecurity groundwork a well-run firm should already have, and the reason the same tools also help you is that a small, organised South African firm can now compete with far larger ones on analysis, documentation and customer response.

What should a business do in the next 90 days?

The next 90 days should go on foundations rather than on picking a model, because every model will change twice before the year is out and the foundations will not. In order:

1. Fix identity first. Enforce multi-factor authentication for every account and turn on conditional access. Agents act with the permissions of the account they run under, so an unprotected account is now an unprotected robot.
2. Classify and label data. Decide what is public, internal, confidential and personal, and apply labels in Microsoft Purview or Google's data protection tools. Copilot and Gemini respect those labels; without them, an agent treats the board pack and the canteen menu the same way.
3. Write a one-page AI policy. Which tools are approved, on which tier, what may never be entered, who approves an agent being connected to a system, and how incidents are reported. A policy that fits on one page gets read.
4. Move staff to business tiers. Retire consumer accounts. The no-training commitment and the admin controls are the whole point.
5. Pilot one agent on one measurable job. Pick a task with a clear before-and-after, such as first-draft replies to a shared inbox or summarising weekly reports, give the agent read-only access at first, and measure hours saved and errors made for a month before widening it.
6. Log and review. Keep audit logs on, and have someone read a sample of what the agent did each week. This is how you catch the confident mistake before a customer does.
7. Train people on the new failure modes. Staff already know not to click phishing links. They also need to know that an assistant can be fed a malicious instruction hidden in a document or a web page, and that "the AI said so" is not a verification step.

What should a business not do?

Three mistakes come up repeatedly, and all three are avoidable. Do not ban AI outright, because staff will use it on their phones with none of your controls. Do not connect an agent to production systems with a global admin account "just for the pilot", because pilots have a habit of becoming permanent. And do not buy a licence for everyone on day one; buy for the team with the clearest use case, prove the saving in rand, then expand.

Where is AI going next?

The direction over the next few years is assistants that see, hear, remember and act, embedded in the tools a business already pays for, becoming steadily more capable at longer tasks. GPT-6 Astra is the current marker on that line, not the end of it. Whether it or its successor earns the label AGI is a matter of definition; its effect on a company is a matter of preparation.

The preparation is unglamorous: identity, data hygiene, a policy, a measured pilot and people who know what the new tools can and cannot be trusted with. Those are the same foundations that make a business secure and compliant today, which is why AMEA treats AI adoption as an extension of security and policy work rather than a separate project. Get the foundations right and every model release becomes an upgrade instead of a scramble.