GPT-6.2 Astra demonstrated strong capabilities in individual tests but failed to deliver consistently in a complex project test: one game-development task required roughly two five-hour cycles and still produced visible lag, missing features and broken asset references.

We notice that the most important takeaway from these hands-on tests is not another set of model records, but the continuing gap between capability, cost and reliability as large models move from “completing one example” to “owning a project.”

What this is

GPT-6.2 Astra is OpenAI’s new model for complex generation and development tasks. The original tests covered image generation inspired by Along the River During the Qingming Festival, a 3D garage recreation, a mechanical watch page, an interactive game controller and a browser game inspired by PUBG: Battlegrounds.

The results for individual tasks were generally strong: the 3D character and vehicle details were relatively complete, the SVG code used to render the controller interface—a code format for describing graphics, icons and illustrations—received high ratings, and the mechanical watch page took about 4 minutes 55 seconds, the controller about 5 minutes 2 seconds, and the image about 6 minutes 24 seconds. The complex game took more than 30 minutes, but its completion level fell short of Claude’s Opus 5.5 and Fable 5.2.

Consumption is even more noteworthy. The tester said that a simple question that required six minutes of thinking consumed about 7% of the package allowance. Other individual examples typically consumed around 10%, while the complex game ran close to two five-hour cycles. Faster performance does not necessarily mean lower usage costs.

Industry view

These tests support one conclusion: models are narrowing the gap between “visual generation and web implementation” and “deliverable software.” Completion speed for individual tasks, coding ability and multi-agent collaboration—where multiple models divide and handle subtasks—are all improving. Enterprises can use these gains to reassess the return on investment for prototyping, internal tools and marketing-content automation.

But the counterargument is equally clear. Complex projects driven by broad instructions continue to expose problems including lag, abrupt viewpoint changes, missing hit feedback, shallow gameplay and distorted external textures. The original test also notes that this was more of an individual, non-standardized experiment with a limited sample and no unified benchmark. It therefore cannot establish that Astra is comprehensively ahead of Claude.

We are more inclined to view it as evidence that “the upper bound of capability continues to rise,” not that “project delivery has become reliable.” What enterprises actually buy is a result that launches on schedule, requires less rework and remains maintainable—not a beautiful page generated for a demo.

Impact on regular people

For enterprise IT: AI coding is already suitable for interface sketches, prototypes and repetitive pages. But business systems involving performance, permissions and stability still require engineers to define acceptance criteria and retain testing and maintenance stages.

For individual professionals: People who know how to write prompts and describe requirements will find it easier to turn AI into an execution tool. The real value lies in decomposing tasks, supplying the right materials and judging the output—not handing over an entire job with one click.

For the consumer market: The barrier to developing independent software and creative services may continue to fall, leading to more custom applications and personalized content. At the same time, template-based products will face stronger commoditization pressure, forcing quality providers to compete more on aesthetics, reliability and after-sales service.