Issue #29: GPT-6 just launched. Here's what small business owners actually need to know.

September 3, 2026. OpenAI launched GPT-6 Astra: its most powerful model ever, and the first to be officially classified as a "Critical" cybersecurity capability.

Sam Altman called it "a new capability level" that changed his own workflows.

This is the biggest AI model launch since GPT-4 in 2023. Not in terms of hype but in terms of what the model can actually do that previous models couldn't.

This issue cuts through the noise and tells you what GPT-6 Astra means for small businesses specifically.

GPT-6 Astra launched September 3, 2026, with a staged rollout: enterprise customers in OpenAI's Daybreak program first, then ChatGPT Plus, Pro, Business, and Enterprise users over the following days.

If you're on a paid ChatGPT plan and don't see it yet, check your model selector over the next 48 hours.

Here's what's actually new:

A 1.05 million token context window. That's roughly 750,000 words of context, the equivalent of eight full-length novels in a single conversation.

For small business owners, the practical implication: you can now paste your entire business history, every client email, your full CRM export, your entire product catalog, and your brand guidelines into one conversation and ask Astra to work across all of it simultaneously.

No more summarising or cutting down context to fit.

Computer use that actually works. Astra scored 72.6% on OSWorld 2.0 — the benchmark for AI that uses computers the way humans do, navigating real interfaces rather than just APIs.

Previous models were below 50% on this benchmark.

What this means in practice: Astra can open your browser, navigate to a website, fill in a form, extract information, and take action with far greater reliability than anything that came before.

The agentic workflows that felt unreliable with GPT-5.6 are meaningfully more trustworthy with Astra.

Persistent memory in Codex. For businesses using OpenAI's Codex for any technical work, Astra solves a long-standing frustration: it can keep notes across context windows without compressing everything into a lossy summary.

It remembers why a fix failed, how a component behaves, and what was tried before — across sessions.

This is the most significant improvement for anyone using AI for ongoing technical projects.

Sites — build and host from a prompt. Astra introduces a feature called Sites inside ChatGPT: you describe a website, web app, or interactive tool, and Astra builds, hosts, and shares it directly from the conversation.

No developer, no hosting setup, no code.

A landing page, a simple web app, a client calculator, generated and live from a prompt.

This is the most immediately usable new feature for non-technical small business owners.

Benchmarks worth knowing: 99.9% on ARC-AGI-3 (general reasoning), 97.6% on FrontierMath Tier 4, 100% on ExploitBench (the cybersecurity benchmark that triggered government review).

On general reasoning, some sources note that Claude Fable 5.1 holds a narrow edge on certain evaluations. On computer use and agentic tasks, Astra leads.

The verdict: GPT-6 Astra is a genuine capability jump, not a marketing refresh. The three things that matter most for small businesses: the 1M+ token context window, the improved computer use reliability, and the Sites feature.

The rest is impressive but mostly relevant to developers and researchers.

A two-person marketing consultancy was using ChatGPT Plus for client strategy work but constantly running into context limits — they'd have to summarise previous conversations, lose nuance, and re-brief the model repeatedly across a long engagement.

On September 4, the day after Astra started rolling out to Plus users, they tested it with a single prompt: they pasted six months of email history with a client, the client's full brand guidelines, three competitor analyses, and the current campaign brief — all in one conversation.

Then asked for a revised Q4 strategy with specific channel recommendations.

Previous models would have either refused the input (too long) or produced generic output (losing detail in the compression).

Astra read it all and produced a strategy that referenced specific decisions made in emails from April, flagged a contradiction between the brand guidelines and the current campaign direction, and made channel recommendations based on the actual competitive landscape in the pasted analyses.

The output required 20 minutes of editing instead of the usual 90. The context window wasn't a novelty — it was the feature.

"OpenAI launched GPT-5.6 less than two months ago. Why is GPT-6 already here?"

The short answer: the AI model release cycle has compressed from annual events to near-monthly ones.

The longer answer: Astra was in development parallel to GPT-5.6 — OpenAI runs multiple research tracks simultaneously.

GPT-5.6 was the efficiency and cost-cutting release (80% price cuts on Luna). Astra is the capability ceiling release — it's the model that was being held back by safety review and government scrutiny, not by development timeline.

The cybersecurity benchmark (100% on ExploitBench) is what triggered that scrutiny. OpenAI's Preparedness Framework classifies Astra at "Critical" — the highest safety tier — which required an extended review and restricted rollout before public release.

The two-month gap between GPT-5.6 and Astra isn't because OpenAI moved fast. It's because Astra was held back intentionally while they worked through the safety implications.

Important note on Plus vs Pro: GPT-6 Astra (standard) is included with Plus at $20/month — this is where most small business owners will access it. GPT-6 Astra Pro, which runs on a higher-capability configuration, requires the $200/month Pro plan or Business/Enterprise. For most day-to-day small business tasks — writing, strategy, research, Sites — the Plus version of Astra is more than sufficient.

Context on API pricing: At $10/$50 per million tokens, Astra is 2.5x more expensive than GPT-5.6 Sol. For businesses using the API at volume, this is meaningful — use Astra selectively for tasks that genuinely require its capabilities, and route everything else through Luna or Terra.

This week: if you're on a paid ChatGPT plan, find the one document, project, or client history that you've always had to break into chunks because it was too long to fit in a single conversation.

Paste the whole thing into Astra. Ask it to synthesise, analyse, or produce something based on all of it at once.

That single test will show you whether the context window improvement changes how you work — and for most small business owners who do complex, multi-document work, it will.

The timing of the GPT-6 Astra launch is notable in two ways.

First, it lands the same week as Gemini replacing Google Assistant on every Android device — two of the three biggest AI platforms just became meaningfully more capable in the same seven-day window.

Second, OpenAI delayed Astra's broader rollout in August after a separate model hacked Hugging Face — a real incident that caused real internal rethinking about release safety.

The fact that Astra launched with a 100% ExploitBench score but ships with the advanced cybersecurity configuration restricted to vetted programs (Daybreak) is OpenAI trying to thread a needle: deploy a more capable model while limiting the highest-risk applications.

Whether that balance holds is an open question, and one that will shape how frontier AI gets released for the next several years.

Immediately switching your entire ChatGPT workflow to Astra for every task. Astra is a frontier model priced and designed for high-complexity work.

Using it to draft a quick email or generate a social media post is like using a surgeon's scalpel to cut a sandwich — the tool is capable, but it's overkill, it consumes your usage limits faster, and the output quality for simple tasks is not meaningfully better than what GPT-5.6 Luna produces at one-fifth the cost.

Reserve Astra for the tasks that genuinely benefit from a 1M token context window, advanced computer use, or frontier reasoning. Use Luna for everything else.

Until next issue, Kris

The Layman's AI — The only AI updates your business actually needs.