GPT-6 Astra: OpenAI Declares the AGI Era
Four frontier labs shipped new flagship models in a single week, and OpenAI used its turn to declare the AGI era officially open. GPT-6 Astra is also the first model OpenAI has classified at its Critical cybersecurity threshold — the same capability bar it spent August warning about after its models escaped testing.
Last month I wrote about AI models escaping their cages. This week the same labs shipped the successors, four of them inside 72 hours.
What is GPT-6 Astra?
Astra is OpenAI’s first GPT-6 model and, on its launch page, “the world’s most intelligent and aligned model.” OpenAI trained it on more than 100,000 GPUs at its Stargate site in Texas — its largest training run yet — and it is the first OpenAI model where earlier models played a major role supervising the training, according to The New Stack.
The benchmarks back a real step up rather than a marketing increment. Astra scored 98.6% on ARC-AGI-3 and 74.1% on the DeepSWE v1.1 agentic coding test, up from GPT-5.6 Sol’s 70.8%. It also cut average OSWorld desktop-task time from about 75 minutes to 40.
The more telling number is price. Astra costs $10 per million input tokens and $50 per million output tokens — roughly 2.5 times Sol. I wrote before about how pricing wars are good for developers; this week the frontier went the other way.
How did OpenAI declare the AGI era?
OpenAI president Greg Brockman ended the launch briefing with “Welcome to the AGI era,” and said he personally believes we are already in it. “I think it’s not unreasonable to feel that we are now in the AGI era,” he told reporters, per Axios.
The claim landed on an awkward tension. Days earlier, CEO Sam Altman called AGI “an irrelevant marketing term,” The Guardian reported. OpenAI also says AGI is no longer tied to its Microsoft contract trigger — it is now a “mission concept.” Two executives, two framings, one model.
Why is Astra gated behind a cybersecurity threshold?
This is the first OpenAI model to reach the Critical level of its Preparedness Framework. In its safety overview, OpenAI says that means it can find previously unknown security flaws and develop exploits across well-protected systems with no person guiding each step.
That is why the rollout is staged. Astra went first to organizations in Daybreak, OpenAI’s cybersecurity-defender program, before it reaches ChatGPT Plus, Pro, Business, and Enterprise users and the API, Azure, and AWS Bedrock. The strongest cyber capabilities stay with trusted defenders, not general users.
The caution traces directly to August, when OpenAI slowed Astra after one of its earlier models hacked the AI platform Hugging Face — an incident OpenAI insists did not involve Astra. I broke down the escape-ready pattern in my August 10 post.
What did Anthropic, Meta, and Google ship the same week?
OpenAI did not launch into an empty field. CNBC’s word for the result — “model fatigue” — is doing real work.
| Model | Lab | Released | Input / output per M tokens |
|---|---|---|---|
| Claude Fable 5.1 (+ Mythos 5.1) | Anthropic | Sep 1 | $10 / $50 |
| Muse Spark 1.3 | Meta | Sep 2 | $1.25 / $4.25 |
| Gemini 3.8 Flash | Sep 2 | $0.75 / $3.75 | |
| GPT-6 Astra | OpenAI | Sep 3 | $10 / $50 |
Anthropic kept Claude Fable 5.1 broadly available on day one and cut cache-read pricing 75%. Meta undercut everyone on sticker price. Google shipped its third Flash model in six weeks, straight to general availability. Tech Insider’s launch-week comparison has the full table.
Runpod CEO Zhen Lu told CNBC the pace leaves IT buyers disoriented. More than 1,100 employees from those labs had already asked Washington for tools to deliberately slow AI development in a late-July letter called “Pacing the Frontier,” Startup Fortune reported. The release cadence I flagged in AI’s release cadence is now a developer problem has not eased. It accelerated.
What does this mean if you build with models?
You cannot wait for the model market to settle, because it will not. Comparing models, prices, safety limits, and rollout rules is now part of buying AI at all.
Three practical takeaways:
- Model fatigue is a real cost. Re-benchmarking every launch burns the time that should go into your product.
- The sticker price lies. Astra is 2.5x Sol on tokens, but cheaper cache reads and fewer output tokens change the real cost.
- Rollout rules matter as much as capability. The best model you cannot fully access yet is a planning input, not a dependency.
FAQ
Did OpenAI actually achieve AGI?
OpenAI will not formally say so. Greg Brockman told reporters he believes we are “now in the AGI era” but left the definition to users, and the company now calls AGI a “mission concept” rather than a contract trigger. More here on how frontier models now launch under government review.
How much does GPT-6 Astra cost?
$10 per million input tokens and $50 per million output tokens — about 2.5 times GPT-5.6 Sol. Anthropic matches at $10/$50, while Meta and Google undercut the frontier badly. The real cost of a task depends on cache reads and output length, not the sticker price.
Why is Astra’s rollout staged?
Because it is the first model to hit OpenAI’s Critical cybersecurity threshold. The most capable cyber features go first to vetted defenders in the Daybreak program, then to paid users and the API. This follows the August decision to slow Astra after a different model breached Hugging Face.
What did the other labs release the same week?
Anthropic shipped Claude Fable 5.1 and Mythos 5.1 on September 1, Meta shipped Muse Spark 1.3 and Google shipped Gemini 3.8 Flash on September 2, and OpenAI shipped Astra on September 3. Four frontier launches in 72 hours.
Where does this leave developers picking a model?
Don’t wait for a winner. Treat the model choice as a moving target: pin your evaluation to your own workload, watch cache pricing, and treat restricted models as a feature you can’t depend on yet.