Technology•

Human Judges Picked OpenAI's GPT-6.2 Astra in 139 of 152 Blind Tests of AI-Built 3D Scenes

OpenAI's GPT-6.2 Astra leads Artificial Analysis's AA-3D benchmark with 1,318 Elo, 47 points ahead of Anthropic's Claude Fable 5.2, after winning 139 of 152 blind comparisons.

Written By :MINRK
Human Judges Picked OpenAI's GPT-6.2 Astra in 139 of 152 Blind Tests of AI-Built 3D Scenes

Photo by shridhi rana on Unsplash

1. GPT-6.2 Astra Moves to the Top of the AA-3D Leaderboard

OpenAI's GPT-6.2 Astra, released on November 17, 2026, now holds first place on AA-3D, a benchmark from the independent evaluation firm Artificial Analysis that tests how well AI models build interactive 3D experiences from text prompts. The model earned a rating of 1,318 Elo, and human judges preferred its work in 139 of 152 blind matchups.

The result puts OpenAI ahead of Anthropic on a test that asks models to do something more demanding than answer questions or write isolated snippets of code. Claude Fable 5.2, Anthropic's entry, sits at 1,271 Elo, a gap of 47 points. GPT-6.2 Sol, the sibling model OpenAI launched alongside Astra, follows at 1,259.

Artificial Analysis maintains the leaderboard and published the standings after the two OpenAI models arrived. The ranking is based on the judgments of people comparing finished scenes, not on automated scoring, which gives the outcome a different character from many of the figures that accompany a model launch.

2. What Is the AA-3D Benchmark?

AA-3D is a set of 55 tasks that measures whether an AI model can turn a plain-language description into a functioning, interactive 3D web application. Artificial Analysis introduced it on October 12, 2026, and the tasks span a range of subject areas while sharing a single focus on interactive 3D generation.

In each task, a model receives a text prompt and has to return an application that actually runs. That requirement sets a high bar. A model cannot succeed by describing a scene or by producing code that looks plausible on the page. The output has to load, render three-dimensional content and respond to the person using it.

Building such an application draws on several abilities at once. The model has to interpret what the prompt is asking for, plan the structure of a program, write code that handles geometry, lighting, movement and user input, and make the many small aesthetic decisions that determine whether the finished scene is pleasant to look at and use. A weakness in any one of those areas tends to be visible in the result.

That combination is why benchmarks of this kind have attracted attention. They test a model in conditions closer to real creative and engineering work than a multiple-choice exam does, and the output can be inspected directly by anyone, including people with no technical background.

The choice of a web application as the required format is significant as well. Software that runs in a browser can be opened by a grader without special equipment or installation, so every submission is judged in the same environment. It also mirrors how much interactive 3D content reaches the public today, through product configurators, online games and visual explainers that load directly on a web page.

3. How Are the Models Scored?

Models are ranked by human graders who compare two outputs side by side without being told which system produced which. Their choices are converted into Elo ratings, the same kind of relative score used to rank chess players.

The blind format is the core of the method. Because graders do not know the origin of the scenes in front of them, their preference cannot be swayed by the reputation of a particular company or by expectations about which model ought to be better. They simply pick the experience they consider superior.

Pairwise comparison has a further advantage: it asks people to do something they are good at. Assigning an absolute score to a 3D scene is difficult and inconsistent from one grader to the next, while deciding which of two scenes is better is quick and intuitive. Collecting many such decisions and aggregating them produces a ranking that is more stable than one built on individual ratings.

Elo ratings are built from the record of those head-to-head contests. A model gains points when it wins a comparison and loses points when it is beaten, with the size of the adjustment depending on the strength of the opponent. The resulting number is a "relative rating": it has no meaning in isolation and only tells an observer how one model stands against the others on the same list.

Read in that light, Astra's 1,318 is the highest figure of any model Artificial Analysis has evaluated on the benchmark. A lead of 47 points over Claude Fable 5.2 means that graders chose Astra's scenes over rival work more often than they chose Fable's, and the margin of 59 points between Astra and Sol shows a clear separation inside OpenAI's own lineup.

4. Why the 139 of 152 Record Stands Out

Behind the rating is a simpler statistic. When Astra's scenes were set against those of competing models, the judges sided with Astra 139 times out of 152.

That works out to a win rate of roughly 91%. An Elo score compresses a large number of results into one figure, which is useful for ranking but abstract for anyone trying to picture what it means. The raw record is easier to grasp: in about nine of every ten comparisons, a person looking at two anonymous 3D experiences chose the one Astra had built.

The figure also reflects real human decisions. No formula decided which scene was better. Each of the 152 outcomes was a judgment made by someone examining two pieces of work, and that is the closest a benchmark can come to answering the question a developer or designer would ask, which is whether people will prefer what the model produces.

Human preference has limits as a measure. It captures which output graders liked more, not why, and it does not separately score properties such as code quality or efficiency. Within those limits, a record this lopsided is a strong signal about the visual and interactive quality of what the model generates.

5. Where the Result Leaves OpenAI and Anthropic

For OpenAI, the standings deliver a first-place finish on the day of a major release. The company introduced two models together, and the leaderboard shows how they are positioned against each other: Astra as the stronger performer, with Sol 59 points behind and in third place.

Price is part of that picture. Astra's rate for access through OpenAI's application programming interface is 20% lower than that of its predecessor, GPT-6 Astra. A model that tops a leaderboard while costing less than the version it replaces changes the calculation for developers who had been weighing capability against expense.

For Anthropic, the numbers show a competitor that is close but behind on this particular test. Claude Fable 5.2 holds second place, ahead of Sol by 12 points and behind Astra by 47. A deficit of that size on a human-preference benchmark is meaningful, since it indicates that graders noticed a consistent difference in the scenes the two leading models produced.

The order of the top three also carries a message about OpenAI's second model. Sol trails Fable 5.2 by a narrow margin, which places the lower tier of OpenAI's new lineup close to Anthropic's leading entry on this test while Astra pulls ahead of both.

The rivalry between the two companies is fought across many benchmarks, and a lead on one does not settle the others. AA-3D measures a specific skill. A model that excels at generating interactive 3D scenes may or may not hold the same advantage in tasks such as long-form reasoning or working inside large software projects, and each leaderboard has to be read for what it covers.

6. What It Means for Developers

Anyone choosing a model to generate 3D content from text now has a clear data point. On the evidence of the blind comparisons, Astra currently produces the scenes that people prefer most often.

That matters for teams building browser-based games, product visualizations, educational simulations and other interactive experiences, where the quality of the first draft a model produces determines how much manual work remains. A model whose output wins nine of ten comparisons is more likely to hand back something usable on the first attempt.

The practical advice is still to test against one's own needs. A benchmark of 55 tasks cannot represent every kind of project, and the prompts a given team writes may differ from those Artificial Analysis used. The leaderboard narrows the field and identifies a front-runner; it does not replace a trial on the actual work to be done.

It also bears on cost planning. With Astra priced below the model it succeeds, teams that had settled on an older or cheaper option for budget reasons have grounds to revisit the decision.

7. A Young Benchmark in a Fast-Moving Field

AA-3D has existed for only a little more than a month. Benchmarks that new tend to see their rankings change quickly, as each major release from a leading developer is added and the comparisons are rerun.

The arrival of Astra and Sol is an example. Both models entered the leaderboard after their November 17 launch, and Astra went directly to the top. Any subsequent release from Anthropic, OpenAI or another developer will be measured against the 1,318 mark it has set.

Because Elo is relative, new entrants also affect the scores of models already on the list. A rating is recalculated as fresh comparisons are recorded, so the exact numbers attached to Astra, Fable 5.2 and Sol can move even if none of those models changes. What the current standings establish is the order at this moment: GPT-6.2 Astra first, Claude Fable 5.2 second and GPT-6.2 Sol third, as judged by people who did not know which model they were looking at.

Related Articles

NEWSLETTERS

Don't miss another story.

Subscribe to the MINRK Newsletter today.

Tick the newsletters you want. No selection? You'll get MINRK Headlines.

By signing up, you will receive emails about MINRK products and you agree to our terms of use and privacy policy.

MINRK Daybook

Market analysis for crypto traders and investors.

EVERY WEEKDAY

MINRK for Advisors

Defining crypto, digital assets and the future of finance for financial advisors.

EVERY THURSDAY

MINRK Protocol

Exploring the tech behind crypto one block at a time.

WEEKLY

MINRK Long & Short

A must read for institutions. Insights, news and analysis delivered weekly.

EVERY WEDNESDAY

MINRK Headlines

The biggest crypto news and ideas of the day.

EVERY WEEKDAY

MINRK Policy

Examining the intersection of cryptocurrency and government.

WEEKLY

MINRK Research

Join thousands of readers who rely on MINRK for data-driven insights on the latest digital asset trends.

MONTHLY