In April 2026, Anthropic announced Claude Mythos Preview. The model had found thousands of zero-day vulnerabilities across every major operating system and browser, including a 27-year-old denial-of-service bug in OpenBSD that had survived decades of human review, and a 16-year-old vulnerability in FFmpeg that five million automated test runs had failed to catch. The headlines treated it as a story about the frontier.
Within days, a security research outfit called AISLE published something that should have been the bigger story. They took the specific vulnerabilities Anthropic had showcased, isolated the relevant code, and ran them through small open-weights models. Eight out of eight recovered substantially the same analysis. One of them ran on 3.6 billion active parameters and cost 11 cents per million tokens. The capability had been sitting in publicly available models for months. The 27-year-old bug was not waiting for the frontier. It was waiting for someone to point a competent model at it with the right scaffolding.
A second instance of the same pattern arrived weeks later, in a different domain. On 20 May 2026, OpenAI announced that an internal reasoning model had disproved Erdős’s unit distance conjecture, an 80-year-old open problem in combinatorial geometry. Within days Anthropic confirmed that the same Mythos model used for the vulnerability work, which had been available internally for roughly five weeks, could solve the problem independently with a shorter and more elegant proof. The capability had been sitting inside their own frontier model for over a month. Nobody had thought to test it on mathematics until OpenAI’s announcement prompted them to try. Sholto Douglas, Anthropic’s head of alignment science, described it publicly as “serious overhang in discoveries.” He used the right word.
That is the shape of the problem. The pace of AI impact is not governed by the pace of AI capability. It is governed by adoption, scaffolding, and organisational absorption, all of which are moving far slower than capability. The gap between what is technically possible and what has actually been deployed is the overhang. And overhangs do not unwind gradually.
What happened to Mythos at the lab scale is happening to AI at every other scale too. Anthropic took five weeks to think of pointing its own frontier model at mathematics. A typical company has not pointed AI at most of the work its own employees do every day. The basic act of taking a capability that exists and aiming it at a problem is the bottleneck, and it is the bottleneck at every layer of the economy at once.
The diverging curves
Capability is sprinting. Cost has done its job. Adoption is barely walking.
METR’s latest data shows the capability doubling time has fallen from seven months to 4.3 months. Inference cost has dropped between 9x and 900x per year depending on task. Yet Eurostat puts enterprise AI use at roughly 20%, only 31% of firms run any AI agent in production, and three-quarters of executives admit their AI strategy is “more for show.”
The constraint has migrated from cost to absorption: workflow redesign, governance, data fragmentation, the slow operational work of moving pilots into production. None of it responds to a 10x price drop by becoming 10x easier. Capability keeps doubling underneath it. The gap widens.
The broken signal
Every technology has had an adoption lag. The internet took years. Cloud took longer. So why is this different.
Because the mechanism that normally forces catch-up is broken. In a healthy market, a competitor deploys slightly ahead of you and your margins start to compress. Customers begin asking why your pricing is what it is. A trade press piece appears. Your sales cycle lengthens by a fortnight. None of it is fatal in any single quarter, but the pressure is continuous, and continuous pressure is the thing that gets boards to move. You feel the future arriving in instalments. That is how laggards traverse a technology gap. Slowly, painfully, with time to adapt.
That is not the regime we are in. Picture every employee in your organisation as a world-class mathematician, applying that level of analytical capability to whatever sits in front of them, at roughly a hundredth of the cost of a human. That is close to what is technically possible with models that exist today. Almost nobody is operating anywhere near it. Competitors are deploying, but at the margin. The pressure that should be arriving in instalments is not arriving, because the gap between what has been deployed and what is actually possible has barely been touched.
A disruptor does not need to close that gap to be devastating. They need to close 5% of it. The absolute capability difference between a firm that has taken even a small step toward the limit and one that has not is not a margin you can compete away with effort. It is a step change. The deployment events that will happen, when they happen, are too large in absolute terms to arrive as instalments. The market is quiet, and the quietness reads like safety. It isn’t. It is the absence of signal in a system where signal is supposed to be how you find out you are about to be in trouble.
The denial pattern shows up clearly in the executive psyche. Bain found that nearly half of tech leaders rated disruptive threats as mild and only 5% as severe. Deloitte’s 2026 data shows companies feel less prepared on infrastructure, data and talent than they did the year before. Three sentences from a Board of Innovation piece capture the cognitive output: “Our customers still want humans.” “Regulation will protect us.” “AI isn’t good enough yet.” Everyone is roughly equally underprepared because no one is feeling the pressure that would normally compel them otherwise. In the AI cycle, the transmission medium is part of the overhang.
How dams break
If the market is not transmitting urgency, what eventually does. A single deployment, somewhere in the value chain, that crosses the threshold from internal experiment to production at scale. When that happens, the pressure that has been accumulating behind the dam arrives at every competitor at once.
Block, in February 2026, is the case worth studying. A single announcement. An internal model step-change had reached the point where 65-75% of engineering, design and product headcount could be removed. Not a multi-year automation programme. Not a sector-wide trend that gave competitors time to adjust. A single decision in a single company. Every other fintech now has to respond within months. The signal that would have arrived as quarterly headwinds across the sector arrived overnight as an existential one.
The longer the overhang accumulates, the higher the cliff. Block’s cut is what one company looks like when it acts while sector-wide adoption has been creeping along at single digits. The first mover at any point on the capability surface forces a discontinuous adjustment proportional to the gap they have just crossed. Companies on the wrong side of that adjustment are not failing because they cannot traverse the gap. They are failing because the signal that tells them the gap exists arrives too late for them to start. The capacity to act is there. The pressure to act is absent. That is recoverable if a company understands the mechanism. It is fatal if it does not.
What to take from this
The overhang is real and measurable. METR on the capability side. Eurostat on the adoption side. AISLE on the cyber side. Mythos on the mathematics side. Block on the disruption side. None of this is speculative.
The absence of competitive pressure is not safety. It is suppressed signal, and the longer the suppression lasts, the larger the eventual release. A company that is a laggard in its sector right now is not merely behind. It is catastrophically exposed, and the market is failing to tell it so. The loudest signal it will get about AI impact in its domain is a competitor announcing that the work is already done. By the time that signal arrives, the response window has closed.