GPT-6 Astra overview: Is it the most powerful model on the market? This is what the benchmarks (and the experts) say

Rarely has so much hype been generated in the world of AI (which, to be fair, is quite prone to hype) as with the announced arrival of Astra, OpenAI’s new GPT 6. In fact, in the days leading up to it, various voices had warned that we were looking at a small revolution in the artificial intelligence arms race, and Sam Altman himself had explained that it would represent “a significant step forward in both capabilities and alignment.”

Now, Astra has already begun rolling out, and reactions just keep coming.

The first, of course, came from Sam Altman himself, as soon as the new model launched:

“GPT-6 Astra is here. We hope it powers a new generation of entrepreneurship, scientific discovery, and creation. We believe it is the best model in the world for computer use, professional work, science, programming, cybersecurity, and more. It took us a bit longer to ensure compliance with the safety and alignment standards required for this level of capability, but we believe the wait will have been worth it. It achieves a score of 98% on FrontierMath (Level 4), 99.9% on ARC-AGI 3, and 100% on ExploitBench”.

Overall, the reactions have been highly complimentary from some of the leading AI experts and communicators around the world, such as Spain’s Carlos Santana (@dotCSV), who expressed his amazement at GPT 6’s performance on some of the main benchmarks OpenAI revealed at launch: “Of the benchmarks included, the most impressive is ARC-AGI 3, where GPT-6 appears to completely saturate the benchmark!! The leap compared with GPT-5.6 Sol is going from 7.8% to fully surpassing it. WOW”

In fact, Santana himself is astonished by GPT 6’s abilities in mathematics and reasoning: “The jump in the model’s capabilities is such that, without using chains of reasoning (that is, the instant model), it can solve tasks that would take a human about 30 minutes with 50% effectiveness, jumping from the 4 minutes achieved by Sol 5.6!”

However, Santana had not yet been able to test Astra himself and, as he noted, was relying on the benchmarks published by OpenAI, which you can review here.

That said, some specialists were able to share firsthand impressions.

Allie K. Miller, an entrepreneur and longtime AI specialist, did have early access to GPT 6, and published a lengthy opinion shortly after launch: “Computer and browser use are simply incredible. I am sure that practically any stable workflow performed on a computer can be at least partially done by AI. It is just so, so good…”

In any case, there was also room to slightly temper expectations, explaining that “it failed on my complex wordplay challenges. No model has beaten them yet” and “others tested a lot around 3D and animations. I did not, but they seem very happy with the results. I saw their results and was not as impressed, but that is okay (…) I would probably still turn to Fable 5.1 for ‘mega architect build’-type requests.”

For his part, designer and Pen.dev founder Tom Krcha shared his experience with 3D design in Astra through this tweet and video:

I had an early access to GPT-6 Astra and I can say, everyone in the world now has a 3D designer at their fingertips.
I gave it an image of a house and asked to create it in 3D with all the details including toys, appliances and furniture. This a full 3D model reconstruction in… pic.twitter.com/VuBOPQH0Ht

— Tom Krcha (@tomkrcha) September 3, 2026

Massive capabilities for design, mathematics, and agentic tasks

As you can see, most of the praise GPT 6 has received so far mentions areas such as its incredible ability to solve complex mathematical problems, handle agentic tasks, and, as Krcha pointed out, its capabilities in three-dimensional modeling.

Developers and designers have highlighted its skill in generating 3D environments, CAD code (BenchCAD), and small playable video games. For example, video game startup Playco tested Astra by creating three video game prototypes and says that most of them worked on the first attempt, requiring 50% fewer manual corrections than the previous model.

And then there is also the matter of security.

Astra is the first OpenAI model to reach the critical level in cybersecurity under its Preparedness Framework. According to OpenAI, without production protections it can find unknown vulnerabilities and develop ways to exploit them in well-protected systems. During evaluations, Astra discovered and used two previously unknown zero-day vulnerabilities… which OpenAI is now reporting to those responsible.

In addition:

• ExploitBench: 100% vs. 78.5% for Sol.
• ExploitGym: 42.4% vs. 30.3%.
• SRE-Bench: 88% on the first attempt and 99.2% within four attempts, compared with 55.9% and 68.7% for Sol

But… is it all that positive?

In any case, not all tests place Astra so clearly ahead of its rivals.

One of the sector’s benchmark references is Artificial Analysis, an independent platform that has already published its first evaluations of the model. In its benchmark index, the Intelligence Index v4.1.1, GPT 6 scores 61.2 points, just three-tenths of a point more than GPT-5.6 Sol (60.9) and below its Anthropic rival, Claude Fable 5.1 (65.7).

In other words, Astra nearly ties with its predecessor and comes in five points below Fable 5.1. It also trails the new Muse Spark 1.3, which Meta has just launched. That is a good indication that the extraordinary results published by OpenAI in certain tests do not mean Astra automatically dominates every benchmark available.

Much more efficient in agentic tasks, more expensive for everything else

Even so, Astra does stand out in certain agentic coding tests, especially when its efficiency is taken into account.

In the Artificial Analysis Coding Agent Index, Astra scores 67 points, roughly matching Claude Fable 5 and Opus 5 and coming in slightly behind Fable 5.1. The big difference is that Astra uses approximately one-third of the tokens used by GPT-5.6 Sol at maximum effort, which represents a reduction of close to 70%. Compared with Opus 5, it uses roughly one-fifth as many.

This jump in efficiency helps offset Astra’s significant price increase, with its tokens costing 2.5 times more than Sol’s. In the coding tasks evaluated, Astra has roughly the same cost per task as GPT-5.6 Sol, but scores two points higher on the benchmark. Compared with Claude Fable 5, the gap is even greater: it achieves a similar score at less than half the cost per task.

The situation is very different in the Intelligence Index. Here Astra is also more efficient than Sol, but the improvement is smaller, because it uses approximately 10% fewer output tokens per task. That reduction does not manage to offset its higher price and, as a result, Astra ends up being about 75% more expensive per task than Sol at maximum effort, despite achieving virtually the same score.

Artificial Analysis also detected a major reduction in hallucinations and improvements in long-duration professional work, although it found notable declines in other tests involving professional tasks, scientific programming, customer service, and reasoning over long contexts.

So… when will we be able to try it?

I am afraid that, for now, GPT 6 is not available to everyone. OpenAI began rolling it out yesterday to a limited set of companies and has said that over the next few days it will reach ChatGPT Plus, Pro, Business, and Enterprise, as well as the API and Azure.

You are probably already eager to try it.

Image: ChatGPT