OpenAI's GPT-6.2 Astra Wins 91% of 3D Coding Battles Against All Rivals
GPT-6.2 Astra tops AA-3D at 1318 Elo, with Fable 5.2 at 1271 and Sol at 1259. Humans preferred Astra's 3D scenes in 139 of 152 blind head to head comparisons. Astra also leads AA-Robotics v0.1 at 79, beating Claude Fable 5.2 at 75.

Takeaways
- GPT-6.2 Astra tops AA-3D at 1318 Elo, with Fable 5.2 at 1271 and Sol at 1259.
- Humans preferred Astra's 3D scenes in 139 of 152 blind head to head comparisons.
- Astra also leads AA-Robotics v0.1 at 79, beating Claude Fable 5.2 at 75.
- Priced at $8/$40 per million tokens, 20% cheaper than GPT-6 Astra, with cached input at $0.40.
- Coding Agent Index is flat vs GPT-6 Astra; Sol actually drops 3 points vs GPT-6.1 Sol.
- Hallucination rate remains high: 59% for Astra, 55% for Sol at max effort.
OpenAI’s GPT-6.2 Astra has taken first place on the independent AA-3D benchmark , which evaluates code-generated, interactive 3D experiences created from plain-language prompts. Human graders chose Astra in 139 of 152 blind, head-to-head comparisons against competing models, a 91.4% preference rate. The benchmark’s focus on runnable scenes makes the result relevant to developers building 3D prototypes, configurators, simulations, and visualizations.
| Rank | Model | Elo |
|---|---|---|
| 1 | GPT-6.2 Astra | 1,318 |
| 2 | Claude Fable 5.2 | 1,271 |
| 3 | GPT-6.2 Sol | 1,259 |
Elo expresses relative performance within the tested pool. Astra leads runner-up Claude Fable 5.2 by 47 points, a substantial separation on a benchmark decided entirely through human preference.
How prompts become 3D rankings
AA-3D v1.0 contains 55 tasks across seven domains. Models generate the code and interaction logic for a 3D scene from each text prompt, and reviewers compare pairs of rendered outputs without seeing the model names. The resulting Elo score aggregates those blind votes into a relative ranking.
The benchmark captures preference for the finished interactive experience, including scene composition, prompt adherence, and usable interaction. Production concerns such as maintainability, browser performance, and accessibility require separate evaluation.
Spatial strength carries into robotics
GPT-6.2 Astra also leads AA-Robotics v0.1, a research preview that tests embodied reasoning through language-based tasks. Embodied reasoning covers planning and decision-making for agents operating in physical settings. Astra scored 79, GPT-6.2 Sol scored 78, and Claude Fable 5.2 scored 75.
The broader model report places Astra second in the Artificial Analysis Intelligence Index, behind Claude Fable 5.2. At the highest effort setting, Astra gained five points over GPT-6 Astra and passed Gemini 4 Carbon at 60 and Claude Opus 5.5 at 59.
Hallucination rates stay high
- Factual reliability: At maximum effort, AA-Omniscience reports hallucination rates of 59% for GPT-6.2 Astra and 55% for GPT-6.2 Sol. Lower values are better on this metric.
- Coding agents: In the Artificial Analysis Coding Agent Index, Astra matches GPT-6 Astra, and Sol falls three points from GPT-6.1 Sol. The evaluation uses the Codex harness, a standardized environment for running agent coding tasks.
Production systems using Astra for factual work should add retrieval, citation checks, schema validation, and human review where errors carry significant costs. Automated builds, browser tests, and visual regression checks can provide similar safeguards for generated 3D code.
Token prices fall 20%
OpenAI lists GPT-6.2 Astra at $8 per million input tokens and $40 per million output tokens. Both rates are 20% below GPT-6 Astra. Cached input falls from $1 to $0.40 per million tokens, a 60% reduction.
| Token type | Price per million | Change |
|---|---|---|
| Input | $8 | 20% lower |
| Cached input | $0.40 | Down from $1 |
| Output | $40 | 20% lower |
Cached input pricing applies when the service can reuse unchanged prompt context, reducing costs for repeated system prompts and long shared prefixes. Artificial Analysis measured an average cost of $4 per benchmark task for Astra at maximum effort, compared with $26.54 for Fable. Per-task costs vary with token consumption and reasoning effort, and API billing follows the token rates above.
Where Astra fits in a 3D stack
AA-3D maps most closely to workflows where a model generates the first version of a scene using Three.js, Babylon.js, or raw WebGL. Likely applications include:
- Product configurators and interactive marketing demos
- Game prototypes, level layouts, and physics playgrounds
- CAD and assembly viewers with animated step-by-step instructions
- Scientific, geographic, and spatial data visualizations
- Interactive lessons that let learners manipulate a scene
Production trials should measure browser compatibility, load time, frame rate, accessibility, prompt adherence, and code maintainability. The benchmark indicates which model people preferred in controlled comparisons; project-specific tests determine whether its output meets a team’s runtime and engineering requirements.
Harder prompts are next
Artificial Analysis plans to increase AA-3D’s difficulty because OpenAI and Anthropic models can already complete most of its basic tasks. When several systems handle the same easy prompts, pairwise votes reveal less about difficult interactions and complex scenes.
The current 1,318 Elo score supports a focused conclusion: among the models and tasks in AA-3D v1.0, graders preferred Astra most often. Teams can use that result to select candidates for internal 3D evaluations, then validate them against their own scenes, frameworks, and deployment constraints.
Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.
Read original report

