Nikesh Arora calls OpenAI's 54% efficiency gain a 'good start' but argues consumption based pricing leaves total cost of ownership out of reach for buyers rolling out AI agents that plan, retrieve, and act on their own.
On July 9, Palo Alto Networks CEO Nikesh Arora told CNBC that AI token prices need to fall by as much as 90% before the technology becomes a standard enterprise tool. He drew the line the same day OpenAI released GPT-5.6 Sol, a frontier large language model the company says runs roughly 54% more efficiently on agentic coding workloads than its predecessor.
"Good start," Arora said of the OpenAI release, before pegging enterprise adoption to a much deeper price cut. He expects roughly a 20% reduction in token costs over the next twelve months, per his CNBC interview, and told TechRadar's coverage of the appearance that compounding efficiency gains would then carry the curve toward his 90% target. Twenty percent is a falsifiable near-term forecast. It is also, on Arora's own framing, not nearly enough.
The gap between those two numbers is where enterprise AI deployment actually lives. Arora's wider point, picked up by PYMNTS and The Next Web, is that consumption-based pricing, where customers pay per token rather than per user per month, multiplies cost across agentic workloads. Copilot-era tools issued a handful of model calls per user session. Agentic systems chain calls to plan, retrieve, act, and verify. Efficiency gains reduce the cost per call. They do not, on their own, reduce the number of calls.
So 54% cheaper inference and 20% cheaper tokens still leave the total cost of ownership for a security operations center, a customer service operation, or a coding team well above the line most enterprises will underwrite without board-level scrutiny. Arora put the demand bluntly: "We need to see the pricing for AI come down."
Arora is one buyer, not a market. He runs a public cybersecurity platform whose products ship into enterprise security stacks, which gives him an unusually broad read on deployment-cost data across thousands of customers. It also locates his forecast as a procurement benchmark, not a model-vendor projection. Frontier labs have an incentive to talk price down to widen the market; enterprise CISOs and procurement leads have an incentive to talk it up, because cheaper tokens expand deployment budgets. CryptoBriefing frames the dynamic as an adoption ceiling rather than a feature gap, and the distinction matters for anyone pricing a multi-year AI roadmap.
Pilot an agentic tool on one team and the bill is manageable. Multiply that by every alert a security operations center triages, every ticket a service desk handles, every pull request a coding team reviews, and the token line item moves from a software subscription into a data-center-style variable cost. Until the per-call economics bend against deployment scale, the use cases that justify enterprise procurement remain narrower than vendor demos suggest.
The watch items are concrete and dated. Arora's 20% forecast is now on the record and falsifiable by mid-2027. Model vendors will publish the next round of per-token list prices as new tiers ship, and the trajectory of those list prices, not the headline efficiency benchmarks, will tell procurement teams whether the cost curve is actually bending. Enterprise buyers rolling out agentic AI in 2026 are effectively betting that vendor roadmaps and consumption-based pricing structures both move in the same direction. Arora's 90% benchmark is the line at which that bet no longer requires a leap of faith.