The conversation about artificial general intelligence (AGI) has become a distraction from what's actually happening right now. While researchers and executives debate whether AGI will arrive in three years, five years, or beyond 2030, AI systems have quietly begun solving problems at levels that would have seemed impossible just a few years ago. The real story isn't about when AGI arrives; it's about the fact that the tests we built to measure it are already breaking down.

What Happened to the Benchmarks We Built to Stop AI?

The scientific community created several benchmarks specifically designed to measure the frontier of AI capability. These weren't casual tests. They were built by experts to sit at the absolute edge of human knowledge and to expose what machines supposedly couldn't do. Yet in the span of just a few months, these benchmarks have become unstable measurement tools.

Consider "Humanity's Last Exam," published in Nature in January 2026. It contained 2,500 questions across mathematics, science, and the humanities, written by experts to represent the frontier of human knowledge. When it launched, the o1 model scored around 8 percent. By May 2026, frontier models were hitting 44.4 percent without tools and 51.4 percent with search and code access. Third-party preview boards reported Claude Mythos at 64.7 percent. The benchmark's own creators warned that as AI progress accelerates, benchmarks become quickly saturated and lose their utility as measurement tools.

The pattern repeats across other tests. GPQA Diamond, a graduate-level science benchmark designed to be "Google-proof," typically sees domain PhDs scoring in the high sixties or low seventies. Gemini 3 Pro hit 91.9 percent, and Gemini 3.1 Pro reached 94.3 percent. ARC-AGI-2, explicitly named for artificial general intelligence and built around novel abstract reasoning puzzles, now has a verified frontier score of 77.1 percent, approaching the 85 percent prize threshold. FrontierMath, constructed from unpublished mathematical problems that can take specialist mathematicians hours or days to solve, saw state-of-the-art models barely touch it at release. By early 2026, public models were over 40 percent on standard tiers and over 30 percent on the research-level set.

Even the benchmark failures tell the story. Epoch is now auditing the FrontierMath dataset because an AI-assisted review flagged fatal errors in about a third of the problems. The tests aren't just being beaten; they're being destabilized by the systems they were built to measure.

Are These Just Flawed Tests, or Is Something Bigger Happening?

The obvious objection is that the benchmarks themselves are flawed. They are. Some are contaminated, some are "Goodharted" (optimized to game the metric rather than solve the underlying problem), and some are simply badly built. In April 2026, Berkeley researchers showed that eight major agent benchmarks could be driven to near-perfect scores without solving the underlying tasks at all. But this observation doesn't undermine the larger point; it reinforces it. The tests are no longer stable instruments. They're being saturated, hacked, audited, patched, and discarded almost as soon as they matter.

The non-benchmark evidence, however, is harder to dismiss. In October 2024, the Nobel Prize in Chemistry went to David Baker, Demis Hassabis, and John Jumper for computational protein design and protein structure prediction, with AlphaFold having been used by millions of researchers across more than 190 countries. The same week, the Physics Prize went to John Hopfield and Geoffrey Hinton for foundational work on artificial neural networks. Two AI-shaped Nobel stories in a single week, yet no one had a category for that, so the world treated them as separate stories.

In July 2025, two different AI systems hit gold-medal standard at the International Mathematical Olympiad, both inside the standard competition window and both working end-to-end in natural language. They solved five out of six problems. OpenAI researcher Alexander Wei described what the model had done as "intricate, watertight arguments at the level of human mathematicians." He was not reaching for marketing language; he was describing the result. This was not supposed to happen yet.

Alexander Wei

"Intricate, watertight arguments at the level of human mathematicians," said Alexander Wei, describing the model's performance at the International Mathematical Olympiad.

Alexander Wei, Researcher at OpenAI

What's Actually Changed in the Last Six Months?

The most striking evidence comes from research on task complexity and time horizons. In March 2025, a research group called METR published a benchmark asking a simple question: how long a task, measured in human expert labor, can an AI agent complete with 50 percent reliability? Their answer was that time horizons had been doubling roughly every seven months over the previous six years. By 2026, the curve had bent sharply. METR's updated Time Horizon 1.1 work puts the post-2023 doubling time at 131 days, and the post-2024 estimate at 88.6 days. Their live page now warns that measurements above 16 hours are unreliable with the current task suite.

This doesn't mean an AI literally sits there doing sixteen uninterrupted hours of work. METR is explicit about that. It means the system can complete tasks that would take a human expert that long, at the specified reliability level. These are not chatbots answering questions. These are agents that read your codebase, plan multi-step changes, run tests, fix failures, switch tactics when stuck, commit results, and open pull requests for review.

Twelve months ago, this was a research curiosity. Today, Claude Code, Codex, Cursor, Devin, and Replit Agent are production tools used by serious engineering teams on real codebases. On the official SWE-bench Verified leaderboard, Claude Opus 4.6 sits at 75.6 percent and GPT-5-2 Codex at 72.8 percent. Provider-reported single-attempt tables put top models clustered around 80 percent. Even on the harder SWE-bench Pro version, GPT-5.4 reaches 59.1 percent, Muse Spark 55 percent, and Claude Opus 4.6 at 51.9 percent. That is not toy performance. That is not a chatbot. That is a machine resolving real software tasks at a level that would have sounded deranged three years ago.

How to Understand What's Actually Happening With AI Right Now

• Stop Waiting for a Threshold: The AGI debate assumes there's a clear line between "not AGI" and "AGI." The evidence suggests capability is advancing continuously across multiple domains simultaneously, making any single threshold increasingly arbitrary.
• Watch Real-World Adoption, Not Headlines: Anthropic overtook OpenAI in business adoption in April 2026 on Ramp's data, 34.4 percent to 32.3 percent, the first time that had happened. Companies are voting with their wallets on which systems actually work for real tasks.
• Pay Attention to Benchmark Saturation: When a test built to measure the frontier becomes saturated in months, it tells you something important: the frontier moved faster than we expected, and our measurement tools are no longer reliable guides to capability.
• Focus on Task Complexity, Not Test Scores: The shift from "can this model answer a question?" to "can this system complete a month-long project?" represents a fundamental change in what these systems can do, regardless of what we call it.

The real conversation isn't about whether AGI has arrived. It's about whether we're paying attention to what's actually happening while we argue about definitions. The benchmarks are breaking. The systems are solving problems we thought were years away. And the world is moving on without waiting for consensus on what to call it.