The Perils of the AI Exponential

What It Covers

METER's latest benchmark data shows Claude Opus 4.6 achieving a 14.5-hour agent task horizon, tripling its predecessor in one generation, while Citrini Research's "2028 Global Intelligence Crisis" report triggers widespread investor anxiety about AI-driven economic disruption and mass unemployment across all labor sectors.

Key Questions Answered

• •METER Benchmark Acceleration: Claude Opus 4.6 recorded a 14.5-hour task horizon on METER's agent benchmark, more than tripling Opus 4.5's 4.8-hour result. GPT-5.3 Codex reached 6.5 hours. The implied doubling rate has compressed from seven months historically to approximately six weeks, though METER warns their task set is nearing saturation and results carry significant noise.
• •Benchmark Methodology Clarity: METER's time horizon metric measures task difficulty in human-equivalent completion time, not continuous AI runtime. A task solved by an AI in two minutes but requiring two hours for a human engineer scores as a two-hour horizon. The 50% success threshold means production reliability standards are not being measured — only capability frontier progression across model generations.
• •Software Sector Repricing Signal: Cybersecurity stocks including CrowdStrike, Okta, and Cloudflare dropped 7–9% following Anthropic's Claude Code Security release, despite minimal product overlap. Analysts at Buco Capital argue selling is rational regardless of specific catalysts because paying 25x revenue multiples becomes indefensible when the software landscape shifts this rapidly, signaling a broad valuation reset rather than targeted disruption fears.
• •Claude Code Revenue Trajectory: Anthropic's Claude Code, launched one year ago as a side project, now generates $2.5 billion in ARR and accounts for nearly half of all Anthropic API tool calls. Tracking this concentration matters for enterprise AI strategy: software engineering remains the dominant AI use case by volume, and Anthropic is using Claude Code to develop and upgrade its own models autonomously.
• •OpenAI Cost Structure Deterioration: OpenAI's updated financial projections show inference costs quadrupled in 2025, compressing gross margins from 40% to 33% against a forecast of 46%. Model training costs are projected to reach $65 billion by 2027. Despite forecasting $28.25 billion in 2030 revenue, total cash burn reaches $665 billion over five years, with profitability not expected until 2030.