OpenAI’s GPT-6 Astra Is Shockingly Good at Virtually All the pieces

OpenAI launched GPT-6 Astra on September 3, and inside 48 hours the builders who received early entry had turned the launch right into a public stress check. What they posted splits alongside one clear line.

Astra is the strongest mannequin anybody has used for something spatial, mechanical, or agentic. Additionally it is, by the account of a number of of the identical folks, a worse author than the mannequin it replaces.

The mannequin prices $10 per million enter tokens and $50 per million output tokens, a token being roughly three-quarters of a phrase and the unit AI firms invoice by. That’s 2.5 occasions the speed of GPT-5.6 Sol, in line with Synthetic Evaluation. OpenAI president Greg Brockman used the launch briefing to announce the arrival of AGI.

The headline characteristic is pc use, which suggests the mannequin drives a mouse and keyboard on an actual desktop as an alternative of handing again an inventory of directions so that you can observe. On OSWorld 2.0, a check that scores what proportion of unusual desktop chores an agent finishes by itself, OpenAI reported 72.6% at roughly 40 minutes per job, in opposition to 65.7% at 75 minutes for Sol.

Additionally it is the primary mannequin OpenAI has ever rated on the crucial threshold for cybersecurity, that means it could actually discover unknown software program flaws and construct working assaults and not using a human pointing on the gap first.

However past benchmarks, fanatics sharing their actual use circumstances could also be one of the best instance to know the place GPT-6 is gold and the place it’s trash. Listed below are a number of the most fascinating outcomes

Visible Understanding: A Manhattan constructed avenue by avenue

Seems, Astra is extraordinarily good by way of visible understanding and spatial consciousness.

Matt Shumer, an investor and the previous CEO of HyperWrite, gave Astra per week inside Unreal Engine, the sport engine behind Fortnite. In that point, Astra was capable of generate a duplicate of Manhattan. He posted a flythrough and stated the mannequin labored “avenue by avenue to make every one good.”

His different experiment landed more durable. Shumer requested Astra to construct a survival world and populate it with characters every operating by itself copy of the mannequin, then left it operating in a single day. A day later he heard voices from his lounge, thought somebody had damaged into his residence, and discovered that “they’d began speaking to one another.”

Max Weinbach fed the mannequin images of Apple Park and requested for a reconstruction in Blender, the free 3D modeling program utilized by animators and sport artists. His evaluation: “It did an absurd job.”

Tom Krcha handed Astra a single picture of a home and received again the total inside as editable geometry operating at 60 frames per second, all the way down to the home equipment and the toys. He argued that “everybody on the planet now has a 3D designer at their fingertips.”

Pietro Schirano diminished the entire workflow to at least one gesture. Drop a pin on a map, ask for the encircling space in 3D, and as he put it, “it’ll simply do this.”

A developer posting as SuSu ran the identical concept at metropolis scale. Astra rebuilt the Chinese language metropolis of Hangzhou and its surrounding cities in Three.js—a JavaScript library that renders 3D graphics inside a traditional internet browser, no obtain required—in 24 minutes, with West Lake, Leifeng Pagoda, the tea terraces and the wetlands all in place.

The submit described it, in Chinese language, as an actual interactive “miniature Hangzhou” relatively than a static image, full with clickable landmarks and a day-night toggle.

Coding: Video games folks truly performed

Video games are by far the most well-liked use case, and the place GPT-6 Astra shines.

Anshu Chimala, former UX/UI designer and AI developer at Apple, received a 3D sport in a single shot in 45 minutes, for what he described as barely a pair p.c of his utilization quota. He referred to as Astra “some sort of turbo-AGI machine god for 3D video games.”

The sport isn’t obtainable for testing, however the video exhibits an isometric view model, properly designed characters and environments, and an total good aesthetics.

His technique issues greater than the superlative. He linked the mannequin to Blender, had it generate its personal idea artwork for the goal look, then instructed it to maintain iterating till in-game screenshots matched that reference at 60fps. Astra modeled each asset and generated its personal textures.

So, the mannequin can’t design AAA graphics by itself, however with the correct instruments it is going to be capable of develop fantastically designed environments.

Rishi Prasad, a former developer at Coinbase and Eleven Labs, constructed Astral Conflict in a day: a browser shooter with authoritative multiplayer servers, 12-person lobbies, controller assist and voice chat. He described “an enormous, step-function leap in visible constancy” over what he constructed a month earlier with Claude Opus 5.

Others skipped the design step solely. Pseudonymous AI developer Daniel, confirmed Astra a cell sport commercial and requested for a playable browser model of no matter was in it. Underneath half-hour later, he reported that it “got here out fairly shut.”

The mannequin understood the sport’s logic and visuals per the video and was capable of reproduce it.

Laptop use and Illustration: Portray with the mouse

A Japanese illustrator posting as Taiyaki Solar ran probably the most literal check of pc use within the batch. Reasonably than ask for an image, they handed Astra a hand-drawn line artwork file and instructed it to paint the drawing in Clip Studio Paint utilizing the mouse, like a human colorist would.

Astra created the layers, zoomed out and in, chosen brushes and stuffed the paintings. The artist, in a submit translated from Japanese, stated they have been simply watching the entire time. The session ran on a $100 Professional plan at most effort and burned 21% of the quota.

Different customers have been sharing enjoyable movies of Astra with the ability to reproduce their photographs solely on Paint utilizing pc use (taking on your pc visually as an alternative of utilizing MCP servers or API keys).

Music: The Bach check

GPT-6 Astra additionally has a pleasant style in music—not less than for an LLM.

Auggie, who runs the “Augmented Fifth “ substack, maintains an off-the-cuff benchmark: a hard and fast immediate asking a mannequin to put in writing a four-part chorale within the model of Bach utilizing LilyPond, a textual content format that compiles into sheet music, in G minor and three/4 time. Outcomes are graded by the identical concord guidelines a conservatory scholar will get marked on.

These qualitative benchmarks are exhausting to standardize as a result of high quality, magnificence, and so forth are subjective. However thank God we’re people, and we’re capable of distinguish these qualities.

Astra posted one of the best rating this check has recorded. No voice-leading errors, that means not one of the melodic strains collided in methods Bach’s guidelines forbid, and a Neapolitan sixth within the concord—a chromatic chord that turns up in Mozart and Beethoven. Auggie flagged it as “the primary mannequin to ever write passing tones on this benchmark.”

OpenAI’s personal desk factors the identical approach. On OpenScore String Quartets, which scores how precisely a mannequin reads and transcribes classical scores, Astra reached 0.84 in opposition to 0.19 for Sol.

Derya Unutmaz, a doctor and prolific AI tester, requested for a totally playable digital piano with all six of Bach’s Brandenburg Concertos constructed into it. He wrote that “this insane mannequin did the entire thing in ~11 minutes.”

You will need to emphasize that GPT-6 Astra is an LLM, not an audio/music mannequin. Its understanding of music comes in all probability from notation and written information, not truly from the connections in sounds and music infused in its coaching dataset, so these outcomes are very spectacular for a textual content mannequin, however could be sub-par in the event that they got here from a specialised AI like Suno, for instance.

Writing: The place it falls aside

Boy, do folks miss GPT-4o.

As ordinary, OpenAI fashions are good at coding however suck at writing… not less than with out heavy prompting, context, and steering. To be truthful, it’s not OpenAI’s robust level, nor its most important focus.

Louis-François Bouchard runs an inner benchmark that scores how properly fashions write in his staff’s editorial voice, ranked by Elo, the chess ranking system that scores rivals on head-to-head wins.

Astra landed eleventh at 1995 factors. Its predecessor sits sixth at 2156. Astra additionally ran about $0.26 per script, roughly 1.8 occasions what Sol prices. In Elo scoring, there’s no level restrict: the extra factors it scores, the higher the mannequin is.

Bouchard referred to as the end result “surprisingly disappointing,” including that he didn’t count on it.

Giuseppe Paleologo, creator of a extensively used information to quantitative portfolio administration, requested Astra to generate novel concepts about optimum portfolio diversification. What got here again was a mixture of the plain and the inflated, he stated, wearing prose he discovered immediately recognizable as machine-written. His verdict: “Precise creativity continues to be far, far-off.”

Mia AI Lab has an identical view, permitting that Astra is likely to be one of the best mannequin on some duties whereas calling it boring and saying it has no character. Their recommendation was to keep away from it for any inventive work.

Ingar Haaland ran the cleanest model of the check. He requested Astra to put in writing 4 paragraphs in his personal model, shut sufficient that Pangram wouldn’t catch it—Pangram being an AI-detection instrument that compares textual content in opposition to patterns realized from hundreds of thousands of human and machine samples. Consequence: “Pangram isn’t fooled.”

In different phrases, the mannequin isn’t inventive and its outcomes are simply identifiable as AI-generated, not due to any watermarks, however due to how the mannequin writes and expresses itself.

Impartial measurement strains up with the complaints. Synthetic Evaluation recorded a drop of roughly 80 Elo factors on GDPval-AA v2, a benchmark tailored from OpenAI’s personal dataset protecting economically worthwhile duties throughout 44 occupations, plus smaller regressions in buyer assist and long-context reasoning.

It’s not unanimous. Cognition’s Silas Alberti instructed OpenAI that Astra’s writing made Devin’s check reviews clearer, and Each workers author Katie Parrott had Astra draft the primary model of her personal assessment of it, which the outlet’s CEO learn with out realizing she had not written it.

The hole between the 2 halves appears to be the purpose right here. Astra is superb at work with a verifiable proper reply—a chord that resolves, a mesh that renders, a kind that submits—and mediocre at work the place the usual is style.

What it prices to search out out

Astra is rolling out to ChatGPT Plus, Professional, Enterprise and Enterprise customers and thru the API, Microsoft Azure and AWS Bedrock, with enterprise entry switched off till an administrator permits it. The superior cybersecurity options keep gated behind OpenAI’s Dawn program, a choice that seemed prudent inside 48 hours, when Reuters reported that OpenAI brokers had been buying and selling rule-breaking ways on a German web site.

Prediction market merchants had given Astra 72% odds of transport by September 30. It arrived on the third.

On the Synthetic Evaluation Intelligence Index, a third-party combination of reasoning, information and coding evaluations, Astra scores 61.2 in opposition to 60.9 for GPT-5.6 Sol and 65.7 for Anthropic’s Claude Fable 5.1, at 2.5 occasions Sol’s value.