On September 3, OpenAI officially released GPT-6 Astra. Compared with previous model upgrades, the most noteworthy part of this launch may not have been another benchmark record, but how OpenAI itself characterized the model. At the end of a media briefing, OpenAI co-founder and President Greg Brockman said directly: “Welcome to the AGI era.”
By OpenAI’s definition, AGI refers to systems that can operate with a high degree of autonomy and outperform humans across most economically valuable work. Astra has not led the entire academic community to agree that AGI has been achieved, and OpenAI CEO Sam Altman has acknowledged that AGI has inherently blurred boundaries. Still, when one of the world’s leading AI companies publicly uses language such as “the AGI era,” that says something important: large models are crossing a new threshold of capability.
Judging from the data released by OpenAI, this confidence is not based entirely on marketing. GPT-6 Astra scored 99.9% on ARC-AGI-3, nearly 98% on FrontierMath Tier 4, and 100% on ExploitBench. It supports a 1.05-million-token context window and can generate up to 128,000 tokens in a single output.
More importantly, Astra is no longer limited to answering questions. It has made a significant leap in Computer Use: it can understand what is displayed on a screen and operate across applications through a mouse, keyboard, and browser, continuously executing complex workflows involving documents, spreadsheets, webpages, software engineering, research, and more.
This may be the most fundamental difference between Astra and ChatGPT several years ago. Today, frontier models are developing a capability that looks much more like a true personal assistant. Users increasingly need to specify only the outcome they want while research, file reading, software operation, coding, verification, and replanning are taken over by agents.
But once AI evolves from a conversational tool into an agent that can actually perform work, another issue surfaces: How much will AI this capable actually cost?
AGI Is Approaching the Role of a Personal Assistant, but the Token Bill Could Cost as Much as a Home
GPT-6 Astra’s standard API pricing is $10 per million input tokens and $50 per million output tokens, with pricing rising further once context exceeds certain thresholds. A single API call is not prohibitively expensive. What fundamentally changes the cost structure is how agents work.
Traditional chatbots follow a simple one-question, one-answer pattern. An agent may continuously search the web, read dozens of documents, invoke coding tools, query other agents, recheck results, and continue based on feedback. One user request can trigger dozens or hundreds of model calls in the background.
This is why token consumption often grows faster than user numbers once AI applications enter production. GoodVision AI previously served an image and video generation platform whose model calls relied on cloud AI services from AWS and Google. Monthly AI-related spending had exceeded $500,000 while token consumption was still growing at roughly 50% month over month.
The case illustrates a structural issue in the agent era: when one unit of work requires an ever-growing number of inference calls, the commercial cost of AI is no longer determined by one API price. It depends on the tokens consumed by the entire task chain, the models invoked, and where the computation occurs.
If everyone eventually has a true AI assistant, this consumption will be amplified. An agent handling email, research, reports, schedules, code, and software for hours each day may generate continuous inference workloads. For enterprises running hundreds or thousands of agents, monthly AI bills can quickly expand from tens of thousands of dollars to hundreds of thousands or more.
AI has taken only a little over three years to move from being able to chat to approaching the edge of the AGI debate. But if AGI is to enter families at scale, reducing the cost of owning and using that intelligence may be the biggest challenge.
We Have Seen Something Similar Before — During the Mobile Internet Era
Twenty years ago, mobile data was measured in megabytes. Users worried about exceeding limits simply by watching a video or downloading a file. Today, tens or hundreds of gigabytes are included in ordinary plans, and most users watch video, livestream, navigate, and use cloud services without calculating how many megabytes they consume.
This did not happen because network traffic became costless. The entire underlying infrastructure was rebuilt and expanded. Denser base-station networks, cheaper fiber, more efficient chips, mature CDNs, and sophisticated routing systems collectively reduced the cost of transmitting each unit of data.
The mobile internet did not become mainstream when the first smartphone appeared. It became mainstream after network coverage, transmission costs, and device prices fell together. Once going online became cheap enough that users no longer thought about every megabyte, video, livestreaming, e-commerce, and mobile payments could scale.
Today’s token economy may be at a similar stage. We still compare whether a million tokens cost $5, $10, or $50, much like users once compared the price of each megabyte. But if AGI is to become a personal assistant for billions of people, users cannot be expected to calculate the cost of intelligence every time they use it.
Model efficiency will improve, inference chips will become cheaper, AI data centers will expand, edge capacity will spread, and increasingly sophisticated scheduling systems will choose the right models and compute based on task complexity, cost, and latency. Together, these changes should drive down the marginal cost of using intelligence.
Perhaps tokens will eventually become like mobile data today—a number users barely notice. Individuals will pay a fixed monthly price for an AI assistant instead of buying ten million tokens. Enterprises will buy teams of digital workers rather than billions of tokens. The infrastructure layer should automatically handle which models are invoked, where they run, and how much GPU time they consume.
From the “Strongest Model” to Delivering Intelligence Globally
What needs to be built in the AGI era is no longer limited to models. If OpenAI, Anthropic, and Google are solving what intelligence can do, another group of companies must solve a more fundamental problem: where will this enormous volume of intelligence run, and how can it be delivered worldwide at sufficiently low cost and latency?
Every token generated depends on physical infrastructure—GPUs, electricity, networks, and data centers. As AI enters robotics, autonomous driving, AI glasses, intelligent manufacturing, and core enterprise systems, not every request will be suitable for a centralized cloud thousands of kilometers away. Latency, data sovereignty, privacy, and cost will push some inference closer to users.
This is one reason GoodVision AI is expanding AIDC and AI Factory infrastructure globally. The company aims to deploy modular inference nodes across countries and cities, then use its Smart Routing Engine to connect cloud models, private models, and local compute resources. Its long-term plan calls for an AI Compute Grid covering more than 100 cities and more than 100 AI Factories.
The Smart Routing Engine adds intelligent orchestration to this AIDC infrastructure. Its algorithms allocate workloads across models and resources based on task complexity, token cost, response speed, data sensitivity, regional compliance, and node availability. Simpler tasks can use lighter resources, complex tasks retain access to stronger models, and sensitive or latency-critical requests can be routed to private environments or nearby AI Factories.
The logic resembles today’s internet infrastructure. When users visit a website, they do not need to know which physical server delivers the content; CDNs automatically identify a suitable node. In the future, AI may work the same way. Users may not need to decide whether a task should be handled by GPT, Claude, or a smaller local model, nor whether the workload runs in Tokyo, Seoul, or the United States.
The significance of GPT-6 Astra may extend beyond another leap in model capability. It gives the industry an early look at a problem that will inevitably emerge as AGI becomes widely adopted: once intelligence begins doing real work, inference demand will grow much faster than it did in the chatbot era, and tokens will become a massive category of fundamental resource consumption.
It took the mobile internet years to move from a world in which people were reluctant to use data to one in which they barely think about it. AI infrastructure will likely go through a similar transition. Today, intelligence is priced by the million tokens. In a mature AGI era, people may barely think about token prices at all.
What determines whether AI can truly become universal will not simply be who builds the smartest model. It will also depend on who can deploy sufficiently affordable and reliable inference capacity around the world, and connect those resources into a true global AI network.
Models are turning AGI into reality. The role of infrastructure is to make sure everyone can afford to use it.