AI Releases Are Accelerating. Progress Isn’t Always Keeping Pace.

Major AI models are arriving at a faster clip. One tally by Samuel Hammond puts the gap between major OpenAI and Anthropic releases at roughly 73 days in 2023 and 18 days so far in 2026—though it does not disclose how it defines “major.” Meanwhile, model names change faster than many users can work out what actually changed. The joke about Claude, GPT and Grok all landing on the same day captures the mood better than a roadmap.

But announcement velocity is a poor proxy for technological velocity. A new model may not represent a clean jump in capability. It may be cheaper, faster, tuned for a different reasoning budget, easier to deploy through an API, or better suited to a narrow class of agent workflows. That is not fake progress. Its primary audience is simply shifting—from the person comparing chat responses to the company pricing an automation programme.

AutomationBench makes the shift visible. The benchmark tests end-to-end business tasks across 47 real-world tools. GPT-6 Astra posts the highest score in the published comparison, about 41%, but sits at the expensive end of the chart. GPT-6 Sol reaches 33.2% at $0.27 per task. Claude Opus 5, shown in its maximum-effort setting, scores 26.9% and costs roughly 11 times as much. GPT-6 Luna is materially weaker on score, but costs only a few cents per task.

This is not a laboratory-clean measure of “intelligence”: the chart compares different models at different effort settings. Astra is shown in max mode, Sol in xhigh, and so on. But buyers do not make laboratory-clean choices. They choose between real configurations, real latency, real failure rates and real invoices.

That points to a less glamorous but more consequential story. The market is not optimising only for the best score on a benchmark. It is optimising for the cost of useful work: handling a support ticket, producing a report, moving a lead through a CRM, fixing routine code, reviewing a document. A modest capability gain that cuts the cost of a workflow in half can matter more in deployment than another headline benchmark record.

So it is misleading to say AI progress has simply slowed down. It is also unwise to treat every release as a new rung on the intelligence ladder. Capabilities are still improving, unevenly and task by task. But lower prices, faster inference and more finely segmented model line-ups are bringing those capabilities into jobs where they were previously too expensive or too unreliable. Progress is moving from the demo to the unit economics.

Progress is moving from the demo to the unit economics.

The next cycle’s central question may not be which model is smartest. It may be: at what price, and with what reliability, can an agent replace—or radically accelerate—a specific slice of work? If that answer changes every week, the flood of releases matters. Just do not measure it by the number of new names. Measure it by the number of tasks that have finally become viable on quality, risk and cost.

There is a catch. Companies cannot absorb an endless stream of new models without repeatedly retraining teams, rebuilding integrations and reworking processes. At some point, customers may ask the industry to slow down—not with a post, but with their budgets.