GPT-6 Astra: “The AGI Era” Has Actually Arrived?

GPT-6 Astra: “The AGI Era” Has Actually Arrived?

I have been researching and using AI for various purposes since last 4 years. When OpenAI launched GPT-6 Astra on September 2-3, 2026, I was both excited and doubtful (since the general prediction was at least 20 years of wait for AGI).

President Greg Brockman closed the briefing by declaring the world had entered the “AGI era.” The company called Astra its most intelligent and most aligned model yet, backing it with headline numbers: 100% on ExploitBench, 98% on FrontierMath Tier 4, and a staggering 98.6% on ARC-AGI-3. (If you don’t know it is a benchmark specifically designed to resist pattern-matching and test general reasoning).

Bold claims deserve in depth scrutiny. Within 48 hours, independent evaluators started publishing numbers that told a much more, I would say, modest story. Please note I am not saying that open AI was lying.

The Benchmarks Don’t Match the Marketing

Artificial Analysis, a widely respected independent evaluator, ran Astra through its Intelligence Index, a total of nine separate evaluations. At its highest effort setting, Astra scored 61. That’s not a leap forward; it’s identical to the score of GPT-5.6 Sol, the model it was meant to replace. It also trails Claude Fable 5.1, which scored 66, and Claude Opus 5 at 63. Its better than Grok (60) but we will have to note that Grok is much cheaper (I don’t have scores for Gemini, sorry). If Astra was truly a step toward general intelligence, you’d expect separation from its predecessor. Instead, there’s a flat line.

What bothers me more that the evaluation on ARC-AGI-3 gap, in which OpenAI’s claimed score was 98.6% actually scored 63% when tested by the ARC Prize Foundation (the actual creator of the benchmark). That’s a 36-point difference. The near-perfect number only shows up under a different, non-standard test configuration. That’s not a small discrepancy; it’s the difference between “this model reasons like nothing we’ve seen” and “this model is decent, but not revolutionary.”

On top of that, Astra costs about 2.5 times more to run than Sol, come on Open AI ($10 per million input tokens and $50 per million output tokens, versus Sol’s $4 and $20 really?). Paying more for a model that performs the same on general intelligence is a hard sell, regardless of the branding around it. I am not sure that it’s just a marketing over stretched numbers or something else. I am not saying that the model has less potential but probably not as much as being marketed.

A Pattern of Overstated Claims?

By the way, OpenAI’s public claims have always kept a fair distance from reality. In October 2025, the company had to quietly delete posts claiming GPT-5 had solved ten previously unsolved Erdos problems in mathematics, this left many experts scratching their heads and wondering what happened. It turned out the model had located existing solutions already sitting in the academic literature, not generated new proofs. That history matters here, because it suggests a recurring gap between how OpenAI frames a launch and what survives outside verification.

As I said and to be fair, Astra isn’t without real progress and potential. Its coding agent performance reportedly improved meaningfully, and OpenAI published math research outputs alongside formal verification materials (Lean certificates, that I struggled to get at some point), which is a more transparent, checkable approach than past announcements. The cybersecurity capability is also genuinely notable, Astra is the first model OpenAI has classified as “Critical” under its own risk framework, capable of finding and exploiting previously unknown vulnerabilities without step-by-step human guidance. That’s a real capability jump, even if it’s a concerning one rather than a purely positive headline. However, I still don’t know it can act as ‘the entity’ (people who has seen mission impossible will relate), maybe not.

Why This Isn’t AGI

Artificial General Intelligence implies broad, flexible reasoning that generalizes across domains the way human intelligence does, not just strong performance on a narrow set of benchmarks under favourable test conditions. What the independent data shows is a model that’s incrementally better in specific areas (coding, some math, cybersecurity) while being flat on general reasoning and dramatically more expensive to run.

Calling that “the AGI era” is a marketing choice, not a technical one. The real story of GPT-6 Astra so far is a familiar one in the AI industry: an ambitious launch narrative. It is safe to say that AI is still a little far from ‘thinking’ maybe it’s still relying on recognizing patters. However I must end the blog with the fact that Astra almost certainly does more structured pattern recognition than earlier models; the chain-of-thought approach helps it decompose problems in a way that mimics reasoning. Hopefully (or not) we will get an actual AGI someday that will not turn rogue on us like in movies.

Have a project in mind?

LET'S WORK TOGETHER

Design and Developed by Buffalo Soldiers

© 2026. All rights reserved @ChandanSharma

Books by Chandan Sharma
Read my books at chandansharmabooks.com

Hindi poetry and practical AI guides, available as eBooks you can read on your phone, tablet or computer. Pick a book, get your copy and start reading today.

Browse the books →