OpenAI's GPT-6.2 Astra Takes First Place on AA-3D Benchmark for Interactive 3D Experiences
Key Takeaways
- •GPT-6.2 Astra won 139 of 152 blind head-to-head comparisons judged by human evaluators on the AA-3D benchmark.
- •Astra recorded 1,318 Elo, 47 points ahead of Anthropic's Claude Fable 5.2 at 1,271, while GPT-6.2 Sol scored 1,259.
- •The AA-3D benchmark spans 55 tasks in which models must produce functioning 3D web applications from text prompts, ranked through blind human votes.
- •On the AA-Robotics v0.1 physical reasoning benchmark, Astra scored 79, ahead of Sol at 78 and Claude Fable 5.2 at 75.
- •OpenAI priced GPT-6.2 Astra at $8 per million tokens, 20% below GPT-6 Astra's $40, while GPT-6.2 Sol reportedly costs about 15% of GPT-6 Astra's cost per task.

OpenAI's GPT-6.2 Astra now sits at the top of AA-3D, a benchmark that tests how well AI models build interactive 3D experiences from plain text prompts. When human judges compared its scenes against rival outputs, they picked Astra's work in 139 of 152 matchups.
The ranking comes from Artificial Analysis, the independent evaluation firm that maintains the AA-3D leaderboard, and arrives shortly after OpenAI released GPT-6.2 Astra alongside a sibling model, GPT-6.2 Sol, on November 17, 2026. The result shifts the top spot back to OpenAI on a leaderboard that has traded places between the two labs in recent weeks.
How the scores stack up
GPT-6.2 Astra posted a score of 1,318 Elo on AA-3D, the highest of any model Artificial Analysis evaluated. Elo is a relative rating, a system long used in chess and other competitive games: a model earns points by beating other models in head-to-head contests and loses them in defeats, so the number only carries meaning next to its rivals.
Those rivals trail by a visible margin. Anthropic's Claude Fable 5.2 came in at 1,271, a gap of 47 points behind Astra, while OpenAI's own GPT-6.2 Sol landed at 1,259 Elo.
The methodology matters here. AA-3D relies on human graders who run blind comparisons, meaning they judge scenes without knowing which model produced them, so that brand recognition does not influence the judging.
What AA-3D actually measures
The benchmark spans 55 tasks across a range of domains, all focused on interactive 3D generation. Models receive a text prompt and must produce a working 3D web application in response. Because the required output is a functioning application rather than a static render, the benchmark tests whether a model can carry a prompt through to usable software, with blind human votes then deciding which of those working results people prefer.
Astra's lead also extends beyond pure 3D rendering into physical reasoning tasks. On AA-Robotics v0.1, another Artificial Analysis benchmark, Astra scored 79. Sol followed closely at 78, while Claude Fable 5.2 registered 75.
Pricing and the cost question
OpenAI has set API pricing for GPT-6.2 Astra at $8 per million tokens, which is 20% lower than GPT-6 Astra's $40 pricing.
Sol tells an even sharper cost story. GPT-6.2 Sol reportedly runs at approximately 15% of GPT-6 Astra's cost per task, while scoring 57 on the Artificial Analysis Intelligence Index.
OpenAI therefore now fields a clear two-tier lineup. Astra is the premium performer that tops leaderboards, while Sol is the budget-friendly option that still competes respectably on the same tests. For developers comparing options, that means the model human judges currently prefer most is also the cheaper of the two Astra generations.
Background: the OpenAI and Anthropic rivalry
Artificial Analysis launched the AA-3D benchmark on October 12, 2026, to specifically evaluate AI's progress in creating interactive 3D experiences. The first evaluations in October showed a competitive landscape in which top models from Anthropic and OpenAI were closely matched, with Claude Fable 5.2 initially taking the lead in November updates. The recent results following the launch of GPT-6.2 Astra and Sol have reshaped the standings. Given how quickly the top of this leaderboard has moved since October, upcoming evaluation updates will show whether Anthropic closes the gap or OpenAI extends its lead.
What this means
For developers building 3D web experiences, the practical takeaway is that GPT-6.2 Astra currently produces the scenes humans prefer most often. The 139-of-152 preference record is the stat that matters most here, because it reflects real people choosing one output over another.
For Anthropic, the results set a clear target. A 47-point Elo gap on AA-3D is meaningful, and Claude Fable 5.2's third-place robotics score adds pressure to respond on both visual generation and physical reasoning tasks.
This article is based on CryptoBriefing reporting edited by Diego Almada Lopez.