GPT-6 Astra Released: OpenAI Acknowledges AGI-Level Capability for the First Time

OpenAI has officially released its next-generation model, GPT-6 Astra. According to an analysis by Bilibili creator "阿猪," this marks the first time OpenAI has publicly acknowledged that its model has reached AGI (Artificial General Intelligence) level. This isn't a minor version bump like GPT-5.5 or 5.6 — it's a full new-generation pretrained model. This article breaks down Astra's capabilities, real-world test cases, and how far it still has to go before reaching true superintelligence.

Benchmark Dominance: ARC-AGI Score Approaches Perfect

Astra delivers a sweeping improvement over the previous generation across virtually every benchmark — coding, computer control, mathematics, and scientific research all show massive gains over prior models.

The most striking number comes from the ARC-AGI test — a notoriously difficult evaluation designed specifically to measure general reasoning ability in AI, scored out of 100. The previous GPT-5.6 managed just 7.8 points, while Astra jumped to 99.9. This isn't incremental progress; it's a near-complete paradigm shift in problem-solving capability.

Of course, benchmark scores can be inflated. The real-world cases provided in the original analysis are far more convincing than any leaderboard.

ARC-AGI (Abstraction and Reasoning Corpus for Artificial General Intelligence) was proposed by AI researcher François Chollet in 2019, specifically designed to measure a model's abstract reasoning and generalization abilities. Unlike traditional benchmarks, ARC-AGI tasks are deliberately constructed to minimize reliance on memorization and training data — the problems are straightforward for humans but extremely challenging for AI, since each one requires inferring a new rule from very few examples. This is why it has long been considered the gold standard for evaluating "human-like general reasoning," rather than knowledge retrieval. Prior top models like GPT-4 consistently scored in the single digits to around 30. Astra's reported score of 99.9, if accurate, effectively renders the test obsolete as a discriminative evaluation tool — which itself speaks to the magnitude of this capability leap.

Computer Control and Long-Running Tasks: Performance That Feels Like Science Fiction

The first case involves computer control. In a financial modeling world championship challenge, Astra completed the task entirely through visual perception — "watching" the screen and operating via mouse and keyboard — at four times the speed of the human champion. This means it no longer relies on dedicated API interfaces; it interacts with graphical interfaces directly, just like a person.

The second case demonstrates sustained autonomous operation. Astra ran continuously for a week to build an entire Manhattan in Unreal Engine — zooming in to the level of individual streets and storefronts with full detail. When finer quality was needed, it could first model assets in Blender and then import them into Unreal Engine, allowing users to experience the generated environment in an immersive walkthrough. This near-unlimited autonomous development runtime was beyond the reach of any previous model.

Astra also handles 3D game development with ease. The analysis noted an amusing small bug: repeatedly pressing the spacebar to jump caused the character to start flying — which, in a way, proves that what it generates is a genuinely runnable project, not just a demo.

Office Productivity: Tables and Tax Forms in a Single Sentence

In everyday work scenarios, Astra-generated PowerPoint presentations have been rated as genuinely professional. It can reconstruct a fully editable Excel spreadsheet from a single image; handling finance and tax forms, lookups, and scheduling are now all one-sentence operations.

As the original analysis put it: it works faster than you, and better than you. For large volumes of repetitive office tasks, this kind of practical capability may matter more than any benchmark score.

Higher Price, But Real-World Cost May Actually Be Lower

The performance gains come at a cost. Astra's per-unit price is 2.5 times that of the previous GPT-5.6.

However, according to the analysis, Astra's token efficiency is significantly higher — in real task comparisons, the actual cost often ends up lower than GPT-5.6. In other words, while the per-token price is higher, fewer tokens are needed to complete the same task, so the total cost doesn't necessarily increase. This is worth keeping in mind for real deployments — judging by the list price alone can be misleading.

Weaknesses Remain: Aesthetics, 3D, and Writing

Astra is not without its shortcomings. A user who tested it early noted that its front-end aesthetics and 3D construction capabilities still fall short of Claude-family models (referred to as "CloudFable 5.1" in the original), and it underperforms in learning and mimicking specific writing styles — in plain terms, its writing still has a noticeable "AI feel."

This is a useful reminder: even a self-proclaimed AGI still has clear ceilings in subjective, aesthetic, and style-imitation domains. A single generational leap hasn't unified capability across the board.

Two Missing Pieces Before ASI

From a broader perspective, Astra is still missing two critical pieces on the road to ASI (Artificial Super Intelligence, which exceeds AGI).

The first is model-level self-improvement — the ability to evolve through continuous learning rather than relying on human-driven retraining cycles. The second is the world model direction: the perception and decision-making grounded in the physical world, as studied by Yann LeCun and Fei-Fei Li. These capabilities may be left for GPT-7 and GPT-8 to realize.

Even so, the capabilities Astra puts in the hands of individuals today already feel like science fiction. You could assemble a team of dozens of Astra Agents and, as a single person, pursue goals that once required companies of hundreds or thousands. That may be the most important takeaway from this release: the threshold for what one person can accomplish is being fundamentally redefined.

Note: This article is based on a single-source analysis by Bilibili creator 阿猪. Some technical data and product names reflect the original video's phrasing. Readers are encouraged to cross-reference with official OpenAI announcements.

AGI (Artificial General Intelligence) and ASI (Artificial Super Intelligence) represent two major milestones on the AI development roadmap. AGI is generally defined as an AI system that can match or exceed human performance across all cognitive tasks; ASI goes further, referring to an agent that vastly surpasses the best human capabilities across nearly every domain. Yann LeCun — one of the "Godfathers of Deep Learning" and Meta's Chief AI Scientist — has long argued that current large language model approaches cannot lead to true AGI, identifying the absence of a "World Model" as the core bottleneck. His view is that AI must develop an internal representation of causality and physical laws — much like how infants build intuition through sensorimotor experience — rather than relying solely on statistical fitting of language data. Fei-Fei Li, meanwhile, is advancing AI's understanding of three-dimensional physical reality through her work on "Spatial Intelligence." These two directions represent the critical bridge from language intelligence to embodied intelligence.