Matthew Berman Tests GPT-6 Astra with Games and Browser Tasks

I have had early access to OpenAI’s GPT-6 Astra and have made a collection of demonstrations, hosted on here.now. In my experience it is the best model I have used, although these are selected projects and my assessment is subjective. A seven-biome scene looks at least comparable to my best attempts with GLM 5.3 and Fable 5.1; I cannot confidently rank them from this example alone.

The first game resembles Fall Guys: characters race over obstacles, with sound effects and a dive control. I initially describe it as a one-prompt project, then clarify that the version shown received one additional feedback prompt from Brian. It is playable in the demo, and I finish a run in about 31 seconds.

After Hours is a 3D city rendered with ASCII characters. I say it took about two prompts and show people moving through a rainy environment, a mini-map and a world that continues to populate as I walk. I borrowed the idea from someone else but could not recall the original creator.

For a SimCity-style project, I used a goal workflow that ran for five days. I watched it create individual assets including fire and police stations, a hospital, a nuclear plant and a university; it still had not finished when I stopped it. The visible result is a browser city-building game with moving people and traffic.

The city project initially had frame-rate problems, so I asked it to optimize for the browser and performance. I show roads, residential zoning, funds, population, happiness and emergencies, and the demonstration appears to respond to my actions. This brief inspection is not a full test of performance or game mechanics.

Another prompt produced a slide deck about data centers using Forward Future branding. I also value Astra’s browser control for everyday tasks such as appointment-making or researching items for sale. These examples show particular tasks, not a guarantee that it can reliably complete every transaction.

I asked the agent to perform browser tasks, write its own recording software and add a timer. The examples include drawing a research workflow in Excalidraw and comparing three Pokémon card listings on eBay; I report about 55 seconds for the comparison. The on-screen label calls the video an edited highlight and excludes setup and rehearsals, so the timer should not be read as total preparation time.

Box sponsors this video, and I present its own AI Complex Work Eval comparing GPT-6 Astra with GPT-5.6 Sol. The displayed overall scores are 77% and 74%, a gain of three percentage points on that dataset. This is a sponsor-reported evaluation of enterprise-content tasks, not an independent demonstration of performance on all knowledge work.

The Box chart shows media and entertainment and technology rising from 62% to 77%, legal from 64% to 72%, consumer products from 87% to 90%, and energy from 77% to 86%. The differences are percentage points; the subsets have their own task composition and do not establish universal performance in each industry. Box also promotes its document platform and says Astra will come to Box AI.

Looking across the projects, I notice a familiar AI aesthetic in both writing and design: several independently prompted projects use variations of forest green and flat visual elements. The results still look good, but repeated defaults could make projects resemble one another. I find Astra responsive to direction, so a small prompt adjustment can produce different designs.

A Rubik’s Cube project scrambles and solves the cube and includes controls for colors and atmosphere. I consider the result impressive but say this kind of demo has become familiar. Another game places a needle among a claimed five million pieces of hay; the player searches one by one. I saw that idea elsewhere but cannot credit its origin.

A small planet demo demonstrates a detailed 3D world with a character moving around it. I initially praise the spacing and lack of obvious clipping, then find that the character can pass through brush and a sheep and notice a small issue around the icebergs. I say this particular project came from one prompt.

At the beach, the character changes from walking to swimming; I cannot enter a boat. I also select waypoints and watch the character navigate toward them, including one near a volcano. These are features shown in the short play session, not a comprehensive test of navigation or collision handling.

Commons is a small town simulation with residents, an economy and social indicators. I describe it as a replay of a simulated town and show controls for recording and experiments such as a dry season or a power outage. The interface displays a local simulation and does not by itself establish that a language model is running every resident live.

The 3D liminal-space horror game took a few prompts, rather than one. I asked for a realistic environment, a subtle antagonist and later jump scares, and explore corridors, a nursery and a pool. I say the assets were created with Astra and speculate that the space continues indefinitely; the video does not verify that extent.

I briefly show a small GTA-style project and a manga, then play a Choo Choo Rocket-inspired game. The player places arrows to direct moving mice into a rocket while competitors attempt to redirect them. I prefer these game results to my own Fable 5.1 attempts, but I have not presented a controlled benchmark.

Finally, I show basic websites for the Galaxy Z Fold, an apple orchard and a rubber-duck club. I find some designs flat and adjust dark mode to view one page. The collection reflects my early hands-on impressions, with different prompts, iteration and time spent across projects.