Study the curve. That is what matters. Not the headline, not the press release, not the breathless social media takes. The curve tells you where the world goes next. And the curve just bent.
On April 7, 2026, Anthropic released Claude Mythos Preview. Ten trillion parameters. A 93.9% score on SWE-bench Verified. A 77.8% on SWE-bench Pro. An 82% on Terminal-Bench 2.0. A 97.6% on USAMO 2026. Double-digit leads over Opus 4.6 and GPT-5.4 across every major benchmark. The most capable AI model ever constructed by any organization on Earth.
And Anthropic locked the door.
Claude Mythos Preview scored 93.9% on SWE-bench Verified, 77.8% on SWE-bench Pro, 82% on Terminal-Bench 2.0, and 97.6% on USAMO 2026.
Verified
That decision matters more than the benchmarks. It marks the first time a frontier lab built a general-purpose model and declared it too powerful for public access. Not too expensive. Not too slow. Too powerful. The distinction creates a new category in the AI landscape: restricted superintelligence. A system that exists, that functions, that outperforms everything else, but that the public cannot touch.
Follow the trajectory. In 2023, the debate centered on whether large language models could reason at all. By 2024, the question shifted to how fast they could improve. In 2025, frontier models began finding real vulnerabilities in production code. Now, in April 2026, a single model finds thousands of zero-day vulnerabilities across every major operating system and web browser. One of those bugs sat in OpenBSD for 27 years. Automated testing tools hit the relevant line of code in FFmpeg five million times without catching the flaw. Mythos caught it.
Anthropic reported 30 billion in annualized revenue on the same day it committed 100 million in usage credits for Project Glasswing.
Verified
Biased BipartisansReal-Time, Evidence-Based News Reports
Unlimited access to your personalized investigative reporter agent, sourcing real-time and verified reports on any topic. Your personalized news feed starts here.
Create Free AccountThe rate of change matters more than the current position. Three years ago, no AI could write a working function. Today, one AI can chain kernel exploits to achieve full system compromise. That acceleration has not slowed. Anthropic itself acknowledges that similar capabilities will proliferate within months, not years.
The second-order effects demand attention. Project Glasswing gives twelve major companies access to Mythos for defensive cybersecurity. Amazon, Apple, Google, Microsoft, Nvidia, JPMorgan Chase, CrowdStrike, and others now scan their infrastructure with a tool that no competitor, no government, no university, and no independent researcher can access. The stated purpose is defense. The structural outcome is a new kind of information asymmetry.
“Given the rate of AI progress, it will not be long before such capabilities proliferate, potentially beyond actors who are committed to deploying them safely. -- Newton Cheng, Frontier Red Team Cyber Lead, Anthropic
Consider what this means for the next five years. If restricted superintelligence becomes the norm, the gap between insiders and outsiders widens at machine speed. Twelve companies get early warning on every critical vulnerability. The rest of the software ecosystem learns about those vulnerabilities weeks or months later, through coordinated disclosure. The defenders with access move first. The defenders without access react.
The competitive dynamics reinforce this trajectory. DeepSeek V4 trained a one-trillion parameter model for 5.2 million dollars. Google launched Gemini 3.1 with 2.5x faster processing. Alibaba released Qwen 3.5-Omni. SpaceX acquired xAI for 250 billion dollars. OpenAI raised 122 billion at an 852 billion valuation. The capital flowing into AI development guarantees that Mythos-class capabilities will spread. The question is not whether other labs will reach this level. The question is what the world looks like when five or ten organizations each possess models that can crack open any operating system on the planet.
DeepSeek V4 trained a competitive one-trillion parameter model for 5.2 million dollars, while OpenAI raised 122 billion at an 852 billion valuation.
Verified
Biased BipartisansThink Further on BIPI.
Where seeking the truth is a journey, not a destination.
Learn moreThe pattern emerging here has a name in systems dynamics: a capability trap. The more powerful the tool, the more dangerous its misuse, the tighter the access controls, the greater the advantage for insiders, the stronger the incentive to build the next tool. Each cycle concentrates capability further. Each cycle makes the case for restriction more compelling. Each cycle moves the world further from the open-research ethos that produced these models in the first place.
Anthropic committed 100 million dollars in usage credits and 4 million to open-source security organizations. Those numbers sound large in isolation. Measured against the 30 billion in annualized revenue Anthropic reported on the same day, they represent a rounding error. The gap between the scale of the capability and the scale of the mitigation tells you which direction the system is moving.
The trend line is clear. AI capabilities grow faster than AI governance. The organizations building the most powerful systems also decide who gets to use them. The era of restricted superintelligence has arrived, and the trajectory from here points toward a world where the most consequential technology in human history operates behind access controls set by private companies. That is not a prediction. That is the curve we are already on.
