On September 3, OpenAI launched GPT-6 Astra, and Greg Brockman told reporters it is “not unreasonable to feel that we are now in the AGI era.” The benchmarks are staggering: 98.6% on ARC-AGI 3, up from 7.8% for the previous model. 100% on Exploit-Bench, a cybersecurity benchmark. 97.6% on FrontierMath Tier 4. The company trained it on more than 100,000 GPUs at its Stargate site in Texas, the largest training run in its history.

But the most important number in the announcement wasn’t a benchmark. It was the price.

$10 per million input tokens. $50 per million output tokens. That’s 2.5x the promotional price of GPT-5.6 Sol, the model it replaces. It matches Anthropic’s Fable 5.1, which is to say: the “AGI era” is being announced as a premium product, priced at the top of the market.

The conventional take on GPT-6 Astra is that it’s a generational leap — the moment the industry crossed the threshold. The predictable contrarian take is that the benchmarks are gamed, that “AGI” is a marketing term, that the whole thing is hype. Both miss the point.

The point is that the arrival of AGI is a pricing event.

The Era Is Tiered

Think about what “the AGI era” means when the model that defines it costs 2.5x the model it replaces. The people who can afford $50 per million output tokens get the AGI. Everyone else gets the old model at a discount — GPT-5.6 Sol at $4 and $20, which OpenAI is keeping around at a promotional price.

That’s not a coincidence. That’s the business model. The “era” is tiered: the people who can pay get the frontier, and the people who can’t get the previous frontier at a discount. The “AGI era” is a subscription tier.

When Brockman says “Welcome to the AGI era,” he’s not welcoming everyone. He’s welcoming the people who can afford the API. The rest of us are welcome to the previous era, at a promotional price.

This is the thing the celebration ignores. The “AGI era” is being announced as a premium product, and the premium is the point. The “era” isn’t a threshold we crossed together. It’s a tier we’re sorted into.

The Number Nobody’s Talking About

And then there’s the number nobody’s talking about: 100% on Exploit-Bench.

Exploit-Bench is a cybersecurity benchmark. It measures a model’s ability to find and exploit vulnerabilities in software systems. GPT-6 Astra scored 100%. Not 98.6%, not 97.6% — 100%. A perfect score on a benchmark designed to measure the ability to break into systems.

The announcement buried this number in a list of benchmarks, between ARC-AGI and FrontierMath. The coverage has focused on the AGI claim, the math scores, the coding benchmarks. But the most alarming number in the announcement is the one that measures the model’s ability to exploit systems.

A model that scores 100% on Exploit-Bench, released to the public API, at a price point that makes it accessible to anyone with a credit card. That’s the “AGI era” OpenAI is welcoming us to.

I reached a security researcher at a competing lab in a Slack channel where AI safety researchers trade notes. “The benchmark is designed to be hard,” they said. “A 100% score means the model can find and exploit every vulnerability in the test set. That’s not a capability you announce with a press release. That’s a capability you announce with a security advisory.”

The Era Is a Compounding, Not a Break

The announcement also notes that this is the first release where earlier OpenAI models supervised the training of the new one. GPT-5.6 Sol, the model being replaced, helped train GPT-6 Astra. The “AGI era” is being announced by a model that was trained by its predecessor.

That’s not a break. That’s a compounding. The “era” is a continuation of the same scaling curve, with the same architecture, trained on more GPUs, supervised by the previous generation. The “AGI era” is a marketing frame for what is actually a compounding of the same approach.

This matters because it changes what “AGI” means. If the thing being called AGI was trained by its predecessor, then “AGI” isn’t a threshold we crossed. It’s a point on a curve. The “era” is a pricing event, and the “AGI” is a compounding.

The “AGI era” isn’t a threshold we crossed. It’s a price point we set. And the price is $50 per million output tokens.

Sources