OpenAI releases GPT-6.2 Astra with 3D strengths and lower API costs
The new flagship model leads in 3D experience generation but trails Google's Gemini 4 Carbon in specific spatial reasoning tasks like floor plan mapping.
The short version
- GPT-6.2 Astra scored 1,318 Elo points on the AA-3D benchmark, the highest mark recorded for building interactive 3D experiences.
- Google's Gemini 4 Carbon leads the Blueprint-Bench 2 with a normalized connectivity score of approximately 58.6%.
- OpenAI reduced API costs by 20% for the new model, pricing it at $8 per million input tokens and $40 per million output tokens.
- The model features a 1.05 million token context window for processing large amounts of data like entire codebases or research papers.
Why it matters
For professionals in design and science, Astra offers a more affordable way to generate complex 3D content. However, the mixed benchmark results show that different AI models still excel at different types of spatial logic.
Astra's new capabilities
OpenAI officially released GPT-6.2 Astra on November 17, 2026. The company positioned this new flagship model as a tool for professionals working in specialized fields such as law, science, and design. Alongside Astra, the company also launched a lighter version called GPT-6.2 Sol, which provides fewer capabilities for less demanding tasks.
One of the most significant updates for developers is the reduction in price. OpenAI cut API costs by 20% compared to the previous model, setting the rate at $8 per million input tokens and $40 per million output tokens. These costs apply to the infrastructure that allows developers to build their own applications using OpenAI's technology.
The model also includes a 1.05 million token context window. This refers to the amount of information the model can keep in its working memory at one time. This capacity allows users to feed the system large documents, such as an entire stack of research papers or a complete codebase, for analysis.
Benchmarks and 3D wins
The new model has already begun collecting high marks on various industry tests. As reported by Crypto Briefing, GPT-6.2 Astra achieved a first-place standing on the AA-3D benchmark. This specific test measures how effectively an AI can build interactive 3D experiences, a task that requires complex spatial generation.
Astra earned 1,318 Elo points on this test, which is the highest mark ever recorded for this benchmark. This suggests that the model is particularly well-suited for designers and developers who need to create visual tools. It indicates a strength in generating 3D content that can be interacted with by users.
Beyond 3D generation, the model performed well on agent-focused evaluations. It achieved a success rate of roughly 78% on the Agents’ Last Exam and reached about 77% on Terminal-Bench-Science 0.1. These scores highlight the model's ability to complete multi-step tasks and handle scientific logic.
Spatial reasoning gaps
Despite its success in 3D generation, Astra faces challenges when interpreting real-world physical spaces. Specifically, it struggled on the Blueprint-Bench 2, which evaluates the ability of a model to create a layout of a room based on interior photographs. This highlights a distinction between generating 3D objects and understanding 2D spatial layouts.
Google's Gemini 4 Carbon currently holds the lead on this specific leaderboard. It achieved a normalized connectivity score of approximately 58.6% on the floor plan test. This suggests that Google's model currently has a better grasp of how rooms are connected and laid out based on visual input.
The results from these tests show that the ceiling for spatial reasoning remains relatively low across the entire industry. While Astra is a leader in some areas, it is not the outright winner in every category. The gap between interpreting a photo and building a 3D experience remains a significant technical hurdle.
The competitive landscape
Independent evaluators are providing a nuanced view of where Astra stands compared to its rivals. Artificial Analysis noted that Astra is in a competitive position against models like Claude Fable 5.2. However, the report clarified that Astra is not the winner across all composite intelligence metrics.
The current market shows a fragmented landscape where different labs lead in different niches. OpenAI is currently leading in 3D generation and agent-focused tasks. Meanwhile, Google maintains an edge in specific spatial reasoning tests like floor plan reconstruction.
Anthropic's Claude Fable 5.2 remains a primary competitor in the conversation for overall composite measures. This variety means there is no single model that dominates every category. Users and businesses will likely need to choose specific models based on whether they need 3D content, spatial logic, or general intelligence.
What to watch
- Whether OpenAI releases updates to improve Astra's spatial reasoning scores.
- How the 20% price cut affects the adoption of Astra in professional law and science workflows.
Try it
Check the current standings of Blueprint-Bench 2 to see how different models compare on floor plan reconstruction.
Sources and links
First reported by Crypto Briefing. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.