Abacus AI says Astra Mode builds 3D games by juggling frontier models

Why it matters

Abacus.AI is betting the durable agent business belongs to the router and workflow owner, even when OpenAI and Anthropic supply the core intelligence.

Abacus.AI, led by Bindu Reddy (@bindureddy) and Arvind Sundararajan, says it has launched Astra Mode, a new configuration for building 3D games and applications by assigning different pieces of the job to different AI models.

In a September 5 post on X, Abacus.AI said GPT-6 Astra handles Blender-based 3D work while other tasks are routed to models selected for them. That makes Astra Mode less a single-model product than an orchestration layer: Astra operates the visual software, while Abacus.AI decides which model should handle the surrounding code, application logic and other work.

The product name comes with a wrinkle. Abacus.AI's current agent documentation describes a configuration called Max, rather than Astra Mode. Max sends coding and application-building prompts to Anthropic's Fable 5.1, according to Abacus.AI, while GPT-6 Astra drives browser and desktop automation. The documentation positions Max for complete application builds, large refactors, debugging and long-running computer-use jobs.

That documented split broadly supports the architecture behind the announcement, though it leaves the boundary between Astra Mode and Max unclear. The X post presents Blender and 3D creation as the defining use case. The help page presents Astra as the computer-use component inside a broader premium setting.

Reddy's model-routing bet

The multi-model approach follows the thesis Reddy has pursued since founding Abacus.AI in 2019 with Sundararajan and Siddartha Naidu. Reddy previously ran AI Verticals at Amazon Web Services, where her organization launched Amazon Personalize and Amazon Forecast. Earlier, she led product for Google Apps, including Docs, Spreadsheets and Slides.

Sundararajan brings the systems side of the same bet. Before Abacus.AI, he led autonomy systems work at Uber's autonomous-vehicle division. He previously worked on the Gmail backend during its initial launch and led machine-learning and serving work for AdSense at Google. Reddy and Sundararajan also built Post Intelligence together before Uber acquired it in 2017, according to Abacus.AI's company biography.

Their original pitch centered on giving companies access to sophisticated machine learning without requiring a large internal data-science organization. The interface and model mix have changed considerably since then. Abacus.AI began with automated model development for enterprise use cases such as forecasting, recommendations and anomaly detection. It now sells an agent that writes software, operates browsers and desktops, connects to workplace systems and deploys applications.

Astra Mode extends that strategy into a demanding class of work. A 3D game requires code, assets, spatial reasoning, scene manipulation, testing and repeated interaction with an engine or editor. Routing each stage to a suitable model gives Abacus.AI room to replace individual suppliers as new models arrive, without asking customers to rebuild their workflow around every frontier release.

Astra gives agents hands inside creative software

OpenAI released GPT-6 Astra on September 3, two days before Abacus.AI's post. OpenAI describes Astra as a computer-use model capable of operating software, browsing, coding and handling multi-step professional tasks. Those capabilities are central to the Blender pitch: generating a script is one job, while opening a project, changing a scene, inspecting the result and correcting errors requires the model to work inside the application.

OpenAI has already promoted game development as an Astra use case. Playco is using the model inside Playbot, an AI development environment connected to Unity and Godot. In OpenAI's case study, Playco said Astra produced three themed prototypes from a common foundation and required 50% fewer manual fixes than the previous model. Those figures come from Playco and OpenAI, rather than an independent evaluation, but they show why Abacus.AI moved quickly to package Astra for 3D work.

Abacus.AI is competing with more specialized products across several layers. Game engines are adding their own AI assistants. Prompt-to-game platforms can generate playable browser experiences. 3D generation companies focus on models, textures and other assets. Coding agents can assemble application logic. Abacus.AI's pitch is breadth: one interface coordinates the models and tools needed to carry a project from an instruction to a running application.

The integration is the product

Reddy and Sundararajan are betting that access to frontier models will become interchangeable, while the agent that chooses and controls them retains the customer. Astra Mode makes that case plainly. OpenAI supplies the underlying computer-use capability, Blender supplies the creative environment and another model may write much of the code. Abacus.AI owns the routing, workflow and user relationship.

That position also carries the hard engineering work. The router must recognize which model should receive each task, preserve context as work moves between models and verify that changes inside Blender match the user's request. A demo can show that an agent produced a scene. A useful development product must keep producing correct scenes as projects become larger, files accumulate and instructions conflict.

Investors have already financed Abacus.AI for a long build. Abacus.AI announced a $50 million Series C in October 2021, led by Tiger Global with participation from Coatue, Index Ventures and Alkeon Capital. That round brought its reported funding above $90 million at the time.

Astra Mode shows where that platform is heading five years later. Reddy's original promise was to spare companies from assembling machine-learning systems themselves. The 2026 version asks users to describe a game or application and let Abacus.AI assemble the models, software tools and compute behind it. Blender is the latest proving ground for whether that abstraction can hold.