GPT-6 Released: AGI Threshold Reached, the Boundaries Between Tools Are Dissolving

A Model Launch Called "Epoch-Defining"

OpenAI has thrown out its coy versioning tradition entirely, making the unprecedented move of releasing GPT-6 (reportedly a self-trained model operating at roughly GPT-5.6 capability). Looking back at the GPT lineage, every version has arrived far later than rumored — GPT-5 came a full year behind expectations. This time, the jump from 5 to 6 arrived fast and hit hard.

The rollout strategy itself is worth noting. Due to a staged release, some users have access while others don't. OpenAI made an interesting promise: if you can't access the new model, you'll receive one daily reset card as compensation. This kind of "alpha strike" subsidy has many long-time users pointing out that GPT now offers the best cost-per-value of any model for everyday use. The author notes candidly that at the Pro tier (around $100/month), daily usage costs come out lower than domestic Chinese alternatives like DeepSeek^Ad or Zhipu.

Blurring the Tool Boundary: GPT-6's Most Fundamental Breakthrough

The most disruptive feature of GPT-6 is how thoroughly it blurs the line between software and tools. In the official launch video, a user controls multiple applications, switches between them, and even connects to a home 3D printer — going from "I want to build a rocket" all the way to a finished rocket model — entirely through conversation, with no keyboard, no coding, and no need to leave GPT.

To understand why this matters, you have to look at how models have historically interacted with tools. Early models relied on MCP (Model Context Protocol) — essentially a restricted interface that software companies opened voluntarily, allowing models to call only the functions those companies were willing to expose. Figma, for example, opened partial MCP access to let Codex draw, but locked real end-to-end capability behind Figma Make, a separate paid product. The reason is straightforward: if everything can be routed around the software itself, what does the software company sell?

The old impasse looked like this: if software companies won't open MCP, models can never truly operate software. And for a model to independently understand software is extraordinarily hard — it can't read text hierarchies the way humans do, and is stuck using a clunky approach of "screenshot, OCR, move mouse, screenshot again." Most software has no accessibility layer for screen readers, modules can't be switched by cursor alone, and the cost of training a model on a complex application is simply impractical.

MCP (Model Context Protocol) was proposed and popularized by Anthropic in late 2024. It's a standardized protocol that allows AI models to interact with external tools, databases, and software services through a unified interface — conceptually similar to how USB standardized hardware connectivity. Developers expose specific functions according to the protocol spec, and models can call them without needing custom adapters for every application. But MCP's fundamental limitation is that it's permissioned — software companies only expose what they're willing to expose, and commercially sensitive or monetization-critical features are almost never included. This makes MCP-based AI control inherently constrained by the ecosystem's commercial incentives rather than the model's technical ceiling. Running parallel to MCP is the computer-vision-driven GUI approach (the screenshot-recognize-click loop), represented by Anthropic's Computer Use and various open-source RPA frameworks — but accuracy and efficiency in complex workflows have never been good enough. GPT-6's breakthrough is opening a third path entirely: using the model's own rapid learning and generalization ability to directly understand the operating logic of any interface.

Tool Comprehension That Scores 99 Out of 100

GPT-6 has shattered this impasse. It exhibits a staggering capacity for tool learning — "look once, know how to use it." According to a review by Bilibili creator "神凡老狗," in a test involving building a Mario game, GPT-6 was the only model that — without being asked — went to GitHub to find original assets, reasoning the way a person would: "There's existing stuff online, why would I recreate it from scratch?" The Mario sound effects, textures, and animations it produced were indistinguishable from the originals, and it spontaneously built out a second level.

Blogger "杨博士说AI" reveals another dimension: traditional benchmarks like SWE-bench and model arenas have been effectively saturated — many domains are approaching perfect scores, forcing the opening of new evaluation tracks. On a test specifically designed for AI — an "unknown-rules mini-game" where players must figure out the rules themselves and win within a time limit — GPT-6 scored 99, nearly perfect, while the highest human score was only in the low 30s and other models hovered around 40.

The significance of this result: it proves GPT-6 can do what humans do — quickly understand and use an unfamiliar tool without any manual. This neatly sidesteps the core obstacle of software companies refusing to open their interfaces.

SWE-bench is a software engineering benchmark released by Princeton University researchers in 2023. It requires models to locate and fix real bugs in actual GitHub repositories, and is considered one of the gold standards for measuring AI coding ability. Its special value lies in testing the ability to solve problems in complex, real-world, under-specified environments — not knowledge retrieval from a closed test bank. As top models approached and broke through 60% on SWE-bench, its discriminating power dropped sharply, pushing researchers to develop SWE-bench Verified and new multi-step reasoning benchmarks. The "unknown-rules mini-game" test described here represents a new generation of evaluation thinking: shifting from testing accumulated knowledge to testing meta-cognition and rapid induction — whether a model can, in a completely unfamiliar rule system, build an internal model through observation and trial-and-error, the way a human infant learns language. GPT-6's enormous gap over humans on this test reveals not just a lead in coding or knowledge, but a qualitative shift in cognitive architecture.

Why This Is the AGI Inflection Point

OpenAI's head of product has publicly called this an "AGI moment." The author has a clear-eyed view of how the AGI standard has shifted — by the old Turing Test benchmark, we've been there for a while, but complex capabilities like physical-world understanding and contextual common sense still had gaps. His working definition of AGI: a system that can essentially replace humans across nearly all work functions.

The deeper logic is this: beyond algorithms and compute, the biggest bottleneck to AI self-iteration has always been the vast sea of human tools and accumulated knowledge that couldn't be directly accessed. Engineers had to translate all of that into a language models could understand — an extraordinarily inefficient process. GPT-6 can autonomously use tools, find assets online, and even invoke hardware like 3D printers. That means models are beginning to acquire the foundational conditions for self-iteration and self-optimization.

The author's point is clear: GPT-6 doesn't mean AGI has arrived today. It means that with GPT-6, AGI is now only a matter of time — and that time can no longer be measured on the same scale as human development cycles. The leap from GPT-6 to GPT-7 to GPT-8 may come far faster than anyone imagines.

AGI (Artificial General Intelligence) has no universally accepted rigorous definition in academia, and the standards used by different organizations vary enormously. The Turing Test (proposed in 1950) used the ability to deceive human judges in conversation as its benchmark — modern large language models cleared that bar long ago. OpenAI has internally used "performing at human level on most economically valuable tasks" as an operational definition, breaking it into five tiers: chatbot, reasoner, agent, innovator, and organization-level AI. DeepMind proposed a matrix using "performance" and "generality" as two dimensions to assess AGI progress. The author's definition — "capable of essentially replacing humans across nearly all work" — aligns more closely with an economics and labor-market perspective than the cognitive-capability framing preferred by technical researchers. This definition matters because it anchors the AGI question in observable social impact: when enough white-collar roles show systematic signs of displacement, an "AGI moment" stops being a laboratory milestone and becomes a real event at the level of economic structure.

The One Weak Point: Cost and Compute

GPT-6's biggest problem right now is cost. Within just three days of launch, compute was already strained, with some users' models being silently downgraded to 5.6. The author notes his own monthly subscriptions across various models are approaching $700 or more — JR, Claude, GPT, DeepSeek, Doubao, all running simultaneously. His sense of value has been completely recalibrated: Photoshop at 1,600 RMB per year once felt expensive; now spending over 1,000 RMB a month barely registers.

In actual use, many tasks that GPT-6 could clearly handle better get routed to cheaper models because the token cost is too high to justify. The author's thesis: the real explosion will come when domestic Chinese models catch up to GPT-6 and drive prices down. Looking at the precedent of OPUS releasing in February with Kimi K3 matching it by July — a month ahead of previous expectations — he makes a bold prediction: if the cycle holds at roughly five months, a cheap, full-platform, full-tool-capable domestic AI could arrive before Chinese New Year this year.

The Countdown for White-Collar Work

GPT-6 has also opened vertical modules in law, finance, science, and mathematics — moving well beyond general intelligence. According to recent reports, mathematician Terence Tao used GPT-6 to advance in days a mathematical problem that previously would have taken years, and reportedly remarked that he was "at a loss for words."

The author's reaction is complicated: when AI can reach into mathematics — the last great cathedral of human intellect — and humans are permanently left gazing upward from the ground floor, the challenge to human cognitive identity is real.

He shares a viscerally affecting personal experience. As an offline design instructor with over a decade of teaching, he had a student with weak foundations and persistent problems — by traditional pedagogy, that student would need years of demanding training to break through. But in the three days after GPT-6 launched, that student used AI to build a portfolio that reached the highest level the author had ever seen in a student — including his most gifted ones. Scrolling through years of his proudest student work, he found that today's students "don't need to know how to do any of it, and can replicate it anyway — and often do it better."

All those layers, channels, curves, filters, saturation adjustments — every technical step that existed as a bridge for human hands — may in the future disappear as if they never existed. Users will simply speak their desired result into a text box. This feeling of "something fundamental breaking inside," the author says, is like the moment AlphaGo laid every Go grandmaster flat on the board. White-collar work — especially the knowledge-intensive roles of the past — may have entered a real countdown.

Terence Tao is an Australian-Chinese mathematician, winner of the 2006 Fields Medal, and widely recognized as one of the greatest living mathematicians, working across harmonic analysis, partial differential equations, and number theory. He has been publicly and actively commenting on AI's progress in mathematics for years, his stance shifting from early caution to genuine engagement. Mathematics is regarded by many researchers as the ultimate "litmus test" for AI reasoning, precisely because proofs demand rigorous logical chains that cannot be faked through pattern matching or statistical association — a single error invalidates the entire proof. Over the past two years, AI's progress on math competitions (such as IMO) and formal theorem proving (such as Lean verification) has already unsettled professional mathematicians. If AI can genuinely accelerate the work of human experts in open-ended mathematical research — not just closed benchmark problems — it means AI has begun to touch the boundary between "retrieving existing knowledge" and "creating new knowledge." That distinction carries implications in cognitive science and philosophy far beyond any engineering application.