US enterprises are increasingly choosing the least expensive AI model that can handle the job, rather than automatically opting for the most advanced solution available. The shift is putting pressure on premium pricing as newer models close the capability gap and token costs keep falling, according to the Financial Times.

Buying the cheapest model that gets the job done

The logic is that costly frontier models should be deployed only when necessary, while routine jobs can be handled by inexpensive models that perform the task sufficiently well.

That pattern showed up in Ramp's September 9 AI Index, which tracks AI spending across businesses. Based on August 2026 spending data, 43.8% of US businesses in Ramp's sample purchased Anthropic products, compared with 39.8% for OpenAI. Ramp also pointed to the growing use of cheaper models as a threat to AI companies whose business models rely on enterprises steadily spending more on frontier systems.

The trend is also consistent with findings from the OECD's agentic AI report published in September 2026. Interviews with 25 organizations indicated that cost savings were a common reason for adopting agentic systems — AI that can plan and carry out multi-step tasks with limited supervision — alongside their contribution to higher productivity and their role in addressing labor shortages.

A price war measured in fractions of a penny

The price of AI models is falling rapidly, just a fresh wave of systems is being introduced. According to earlier Cryptopolitan coverage, OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, priced at least 50% below the previous GPT-5.6 models. GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens, with tokens serving as the standard unit of text that models process and on which API pricing is metered.

In the same week, Anthropic launched Opus 5.5 while xAI introduced Grok 4.7, with the debuts landing in a market increasingly driven by price-performance criteria. According to Axios, cheap but still efficient models represent the industry's new competitive race.

The long-term downward trend is even more striking. According to Epoch AI, the price of achieving a constant level of AI performance has fallen by around 47% on a quarterly basis since 2023 — roughly 13 times per year.

Cost per finished task matters more

As price-based pressure intensifies, the metrics by which companies measure AI success are shifting. Benchmark leadership has not diminished in importance, but greater focus is now placed on how much clients pay for effective performance.

Zapier's AutomationBench tests AI agents on end-to-end workflows spanning 47 business tools. As of September 28, Claude Opus 5.5 led the table with 42.47% efficiency at a price of $1.44 per task, while GPT-6 Sol registered 33.2% at XHigh for only $0.27 a task.

The same tradeoff appears on Artificial Analysis. GPT-6 Sol at maximum effort scores 48 on its Intelligence Index at an estimated $1.06 per task, while Xiaomi's open-weight MiMo-V2.6-Pro — a model whose weights are publicly released and can be run outside the vendor's own infrastructure — scores 46 at just $0.13 per task.

That gap is narrow enough that multi-model routing becomes viable. Instead of relying on a single vendor for all workloads, companies can assign the right model based on the relative importance and complexity of each task.

Cheaper tokens can still mean bigger bills

Lower prices do not necessarily translate into lower AI spending. According to Axios, the Jevons paradox — first described by the 19th-century economist William Stanley Jevons, who observed that greater efficiency in using a resource tends to increase its total consumption rather than curb it — indicates that more affordable resources lead to their greater consumption. Gartner offered a similar warning in its August 17 forecast, estimating a more than fivefold rise in the inference cost of agentic workflows by 2028.

Less expensive models can also widen access in other parts of the world. According to Microsoft's Global AI Diffusion Report, affordable and open models can help improve availability, but infrastructure, connectivity, and skills will play a vital role in determining who is able to take advantage of them.

For corporate buyers, the message is becoming stronger: the best AI might not be the smartest AI, but the kind of AI that does its job efficiently at the lowest cost.

Source: Cryptopolitan