AI capability and revenue are compounding in months, not years, and most institutions are not calibrated for it

The number that clarifies the moment belongs to Brendan Foody. On the Apex benchmark, frontier models scored 1% twelve months ago. They now score 40%. That is not a gradual curve. It is a step change compressed into a calendar year, and it lines up with the doubling rate Nathan Labenz describes from the METR chart: task-length capability has been doubling in just under four months, implying an 8x to 12x increase over the course of a single year.

The revenue figures move at a comparable rate. Krishna Rao says Anthropic started the year at roughly $9 billion in annualized run-rate revenue and ended the most recent quarter north of $30 billion. Marc Andreessen puts the trajectory in competitive terms, arguing that Anthropic and OpenAI are currently adding more revenue per month than Meta, Google, or Microsoft. Elad Gil notes that Anthropic and OpenAI each reached $1 billion in revenue in approximately one year. Amjad Masad offers the most compressed single-company data point: Replit went from $2.5 million to $250 million in annual revenue in one year and is tracking toward $1 billion in the current year. His daily ARR jumped from $1 million to $2 million in the two days after Replit Agent launched.

Capability improvement, not just adoption, is driving those numbers. Dylan Patel says Anthropic’s models advanced from L4 to L6 engineer level in two months, and that Mythos represents the biggest step up in model capabilities in approximately two years. Patel also estimates that the economy is currently spending roughly $40 billion on Claude 4.6 Opus-tier models, with that figure potentially reaching $100 billion by year-end. Gavin Baker observes that Claude Opus is now generating 70% fewer tokens for the same question compared to earlier model versions, a finding that squares with what Cat Wu reports: token cost per engineer rises with each model jump, because better models attract more delegation, not less.

We started the year with about $9 billion of run rate revenue and we ended the quarter with, you know, north of $30 billion of run rate revenue. Krishna Rao

The benchmark data reinforces the direction. Carina Hong reports that Axiom Math scored 120 out of 120 on the 2025 Putnam exam, outperforming the best human score of 110 and the best large language model score of 103. The same system achieved 99% on Code Marina, solving 187 of 189 problems with no modification to its architecture. Separately, the system’s proof trees scaled from 40 nodes to 4,000 nodes, a 100-fold increase. In security research, Rao reports that Anthropic’s Mythos model found 250 vulnerabilities in an open-source codebase where a prior model had found only 22.

Scaling laws, the theoretical backbone of these gains, appear intact. Mark Chen notes they have held for almost ten orders of magnitude and sees no reason they should stop. Rao says flatly that from what Anthropic observes, the scaling laws are not slowing down, a direct rebuttal to a widely circulated view that the field had hit a ceiling. Nathan Labenz adds that pre-training never really stopped working, a claim consistent with the observed benchmark trajectory.

The practical consequence is that the relevant unit of time in AI development is no longer the year. Martin Casado puts the model relevance window at three to nine months. Gavriel Cohen makes the operational implication concrete: agents cannot simply be deployed and left on a fixed model version for years the way traditional enterprise software can. The underlying capability changes too fast. Descript’s Laura Burkhauser says her team will have a new model evaluated and in the product within 15 minutes of its release. MiniMax, per Olive Song, ships a new model version approximately every month to a month and a half.

What the evidence collectively describes is a system in which every major variable, capability, cost efficiency, revenue, benchmark ceiling, and deployment speed, is moving in the same direction at the same time, and doing so faster than most institutions assumed was possible. David Heinemeier Hansson says the past three months produced more churn in his mental approach to computers than any prior period in his entire computing life. That is a personal data point, but it echoes across nearly every domain represented here. The question the pace raises is not whether the trend is real. It is whether the planning horizons of organizations that depend on AI, or compete with it, are calibrated to a doubling time measured in months rather than years.