GPT-6 Astra: OpenAI declares the 'AGI era' and reinvents enterprise agents

The day OpenAI said "welcome to the AGI era"

On September 3, 2026, Greg Brockman, President of OpenAI, closed the GPT-6 Astra launch briefing with a line that will echo for years to come: "welcome to the AGI era." He went even further: when asked personally, he stated, "For me personally, I do think we're there." This is not standard corporate rhetoric — it is a philosophical and technical position statement made by the company's second most powerful executive in front of clients, investors, and the press.

For those who follow the sector closely, the provocation was expected. What was not expected — at least not at this pace — was the level of technical evidence the company brought to back up the argument.

What GPT-6 Astra actually is

GPT-6 Astra is the first model in OpenAI's "sixth-generation family," directly succeeding GPT-5.6 Sol. The company describes it as a "generational leap" over its predecessor — and the benchmarks released (all self-reported by OpenAI, worth noting) give substance to that claim.

On ARC-AGI-3, a widely used benchmark for measuring adaptive reasoning, Astra scored 98.6% — compared to 7.8% for GPT-5.6 Sol. The gap is staggering. On FrontierMath Tier 4 v2, it reached 97.6% (versus 87.8% for the predecessor). On GPQA Diamond, 96%. On BenchCAD, 95.9%. On DeepSWE v1.1, which measures software engineering, 74.1%. And on OSWorld 2.0, which evaluates autonomous computer use, 72.6% — versus 65.7% for GPT-5.6 Sol, with approximately 47% less time per task.

The Exploit-Bench, which tests offensive cybersecurity capabilities, recorded a perfect score — 100%. That is precisely the number that made this launch more complex than any before it.

The context window that changes what is possible

Astra operates with a context window of 1,050,000 tokens and a maximum output of 128,000 tokens. In practice, this means a model can ingest entire contracts, complete codebases, CRM conversation histories, quarterly reports — and still have room left to reason. The effective input, when maximum output space is reserved, comes to approximately 922,000 tokens.

Training was conducted on more than 100,000 GPUs at the Stargate complex in Texas — the largest training run in OpenAI's history, according to the company itself. The knowledge cutoff is April 30, 2026.

Why the model launched restricted — and what that says about real risk

GPT-6 Astra is the first OpenAI model classified as "Critical" under its Preparedness Framework for cybersecurity capabilities. This is not a marketing label — it is an internal risk classification that compelled the company to gate access in an unprecedented way.

The initial availability phase, starting September 3, 2026, was restricted to organizations in the Daybreak program — a closed, vetted early-access program. Full cybersecurity capabilities remained tied to Daybreak Blue, aimed exclusively at defense organizations, with selection on a case-by-case basis. Access for Plus, Pro, Business, and Enterprise subscribers and developers via API — including distribution through AWS Bedrock and Microsoft Azure — was announced for the "days following" the launch. As of September 8, 2026, the date of this publication, the model was confirmed as active for Daybreak organizations; the rollout to other tiers was ongoing.

No geographic exclusions were announced — which includes Brazil, Italy, and the other markets where I operate.

What changes in how a company operates

Here is the question that matters for anyone reading this article with a business to run.

Astra was designed to execute end-to-end workflows, without human intervention in the loop. It operates directly inside spreadsheets, presentations, CAD tools, Python notebooks, browsers, and CRM systems. It can fill out forms, update records, manage calendars, search, draft emails and documents, analyze data, and verify website functionality. It can also build and publish websites, web applications, and games directly within ChatGPT.

This is not single-task automation — it is delegation of entire processes. The practical difference: today, an SMB using AI still needs a human to stitch the steps together. With Astra, the model executes the full sequence. Tax return preparation, prototype development, architectural rendering, legal memorandum formatting — all within a single agent, with reasoning adjustable across five levels: low, medium, high, xhigh, and max.

The cost: real, but calculable

Transparency is essential here. GPT-6 Astra is the most expensive mainstream model OpenAI has ever launched. Through the API, standard pricing is $10 per million input tokens and $50 per million output tokens — approximately 2.5 times the price of GPT-5.6 Sol, according to the company itself. Fast mode, which delivers up to 2.5 times more processing speed, costs double: $20 for input and $100 for output per million tokens.

Cached input runs $1 per million tokens in standard mode and $2 in Fast mode — making repetitive workflows significantly cheaper at the margin. For Plus, Pro, Business, and Enterprise subscribers, Astra is included in existing plans, with additional credits available for usage above quota. Enterprise administrators must enable the model manually per workspace — it is off by default.

The figure that anchors the ROI calculation: on the DeepSWE v1.1 benchmark, cost per task via API was approximately 57% lower than GPT-5.6 Sol in each model's best-performing configuration — according to self-reported data from OpenAI. More expensive per token, but more efficient per delivered result.

What I see on the horizon

In June 2026, OpenAI filed documentation with U.S. regulators anticipating a valuation of up to $1 trillion — a figure self-reported in the filing. It is no coincidence that Astra arrives at this moment: it is the product that justifies that thesis before investors, partners, and the market.

For the SMBs I advise in Brazil, Italy, and the U.S., the message is direct: the complexity threshold that justified keeping processes exclusively human has just dropped. Not because AI is perfect — it is not — but because the cost of not experimenting has become greater than the cost of failing fast and adjusting.

GPT-6 Astra is not an incremental update. It is a shift in the underlying premise of what a software system can do autonomously. And companies still in observation mode need to decide, now, whether they will adjust their operations proactively — or react once their competitors have already done so.