The expense of obtaining identical-quality responses from artificial intelligence models has declined roughly 13 times per year since 2023, yet just 37% of survey participants reported that AI contributed anything to their earnings before interest and taxes, according to the State of AI Report 2026. Nathan Benaich and Air Street Capital released the 244-slide ninth edition on October 8, 2026, documenting a striking disconnect: while OpenAI and Anthropic's combined revenue run rate jumped from approximately $30 billion to roughly $105 billion within eight months, most organizations still can't demonstrate financial returns from the same technology driving that growth.
Epoch AI's analysis, published September 22, 2026, measured the cheapest route to reach a fixed benchmark score across five categories spanning mathematics, science, and games. The price fell about 47% each quarter since 2023, with the steepest drops—roughly 66% per quarter—occurring immediately after a model first hit a new top score. Two years after reaching that milestone, the decline slowed to approximately 32% quarterly, or about 4.7 times cheaper annually. One concrete example: in early 2025, an OpenAI reasoning model achieved a set score on GPQA Diamond, a PhD-level science exam, for around 30 cents per question; about 18 months later, GPT-5.6 Luna reached the same score for $0.0004 per question, representing a roughly 725-fold cost reduction. McKinsey's 2026 State of AI survey of 1,719 respondents found that 80% reported greater productivity from AI, but only around 6% qualified as "high performers" who attributed at least 5% of their EBIT to AI and considered the impact significant. When Anthropic released business plugins for Claude Cowork in early February 2026, nearly $285 billion in software market capitalization vanished, and the S&P 500 software and services index dropped 26% from its October 2025 peak, though the sector recovered that value by September 2026.
The report attributes the profit gap primarily to what Benaich calls a "human skill issue" in his launch essay—most organizations haven't learned how to configure and deploy AI effectively. McKinsey's high performers were more likely to redesign workflows around the technology and track whether those changes delivered returns, rather than simply purchasing access to models. Epoch AI warned that most companies will capture only a portion of the theoretical cost savings because labs often optimize models for popular benchmarks, which measure narrower capabilities than real-world usefulness, and because few organizations consistently switch to the cheapest model for each task. Agent systems compound this challenge: a single agent task can trigger dozens of billed model calls for planning and self-checking, so lower per-call pricing can still produce higher total bills. In August 2026, Claude "led" 26% of Anthropic's model research and development work with human supervision, according to Anthropic's R&D Automation Index, up from under 1% six months earlier—though Anthropic noted that when Claude graded its own contributions, those assessments matched staff evaluations only 59% of the time.
The report makes nine predictions for the next 12 months, four of which directly affect ordinary businesses: a major payment network will introduce dispute rules assigning liability for AI agent purchases, a US state will pass legislation requiring companies to accept cancellations and claims from customers' AI agents, an agent will cut its failure rate in half on new tasks after a month of customer use with the same underlying model, and US AI labs will launch frontier cyberdefense products. The deck closes with a single-line prediction reading "AGI 2027" with no supporting detail, and notes that last year's forecasts achieved a 20% strict accuracy rate, with two of ten predictions coming fully true. Benaich writes that if he were running a small team, he'd examine two things before year-end: monthly AI spending and which actual processes changed because of it, because the models will continue getting cheaper on their own but the payoff only materializes once the work itself is redesigned. The clearest beneficiaries so far remain the two labs selling the models—they collect revenue the moment their technology is used, while their customers see profit only after restructuring operations. For software buyers, the volatility matters because the tools they pay for monthly may shift pricing or features as vendors respond to competitive pressure from AI-native alternatives.