Why Palo Alto says AI token prices must fall before AI scales

For enterprise buyers, the AI debate is starting to look less like a benchmark contest and more like a budget review. Palo Alto Networks chief executive Nikesh Arora said AI token prices need to fall about 90 per cent before large organisations can deploy the technology at scale, putting a hard number on the cost problem that has sat behind many corporate pilot projects.

That framing matters because it shifts the conversation away from raw model quality and towards unit economics. In the same week, OpenAI said GPT-5.6 Sol was 54 per cent more token efficient on agentic coding tasks, while Meta launched Muse Spark 1.1 with what Mark Zuckerberg called much more affordable pricing. Capability is still the headline. Cost is becoming the gate.

But the insider view is not identical to the buyer view. Vendors read the same price compression as proof that AI is moving into a scale market, one where lower token costs can unlock much higher usage. The tension sits at the centre of enterprise AI right now: buyers want software they can roll out without blowing through budgets, while model providers need enough revenue to justify the compute bill behind the service.

In Arora’s interview remarks, the Palo Alto boss made clear that recent efficiency gains do not yet solve the deployment problem.

I think 54% is a good start. I think we probably need another turn at it.

— Nikesh Arora, CNBC

AI pricing is becoming a procurement problem

Arora’s number lands because it sounds like procurement language, not founder rhetoric. Large companies already compare cloud products by usage tiers, commit levels and surprise-bill risk. AI models are moving into the same bucket. Once a technology leaves the demo stage and enters workflow software, finance teams want to know what happens when hundreds or thousands of staff use it every day.

That is the analyst concern in plain terms: do lower token costs reduce the bill, or do they simply encourage more usage? TechCrunch argued this week that the AI return-on-investment debate has come back at trillion-dollar scale, and Business Insider reported that Sam Altman told Sun Valley attendees the industry’s central question is how to make AI cheaper. Cheaper inference helps, but it does not guarantee a smaller enterprise spend line if broader deployment turns occasional use into always-on consumption.

Not even close.

A 54 per cent efficiency gain sounds large until it is set against Arora’s 90 per cent benchmark for mass rollout. That gap helps explain why enterprise buyers are starting to treat model selection less like a branding exercise and more like a capacity-planning decision. They are comparing usage caps, latency, workflow fit and the likelihood that a pilot can survive contact with an annual budget cycle.

Inside large IT estates, procurement teams also want guardrails on where premium models are actually worth it. Internal copilots for summarising tickets or drafting boilerplate can move to lower-cost systems faster than regulated code review, security analysis or customer-facing agents that carry higher error costs.

For Australian technology leaders, that logic is familiar. Cloud migrations, cyber tooling and SaaS renewals rarely win approval on technical merit alone; they have to show predictable spend and a credible path to operational value. AI is heading into the same discipline.

Cheaper models do not settle the vendor equation

From the vendor side, however, lower prices are not simply a concession to buyers. They are a bet that demand will expand faster than prices fall. Meta’s new Muse Spark 1.1 launch was pitched as both a performance move and a pricing move, and TechCrunch’s coverage of the release placed it squarely in the fight with OpenAI and Anthropic over coding workloads.

Zuckerberg’s own framing pointed to that volume strategy.

We think that there’s a real ability to be able to offer frontier or very high-level intelligence at a much more affordable cost.

— Mark Zuckerberg, via Business Insider

That line partly answers the insider question about how low prices can go. They can fall a long way if vendors believe lower rates will bring more developer workloads, more enterprise seats and more downstream platform revenue. Put differently, model makers appear willing to sacrifice price per token if the trade-off produces much larger token volumes.

The sceptical reading is harsher. The Register argued this week that AI inference is becoming a bargain-hunter’s market, with only a few luxury models left at the top. CNBC also reported earlier this week that Chinese AI models are gaining ground with US companies as OpenAI and Anthropic costs rise. If that trend holds, premium frontier labs will have a harder time arguing that every workload needs the most expensive option.

Premium AI now needs a clearer reason to exist

This is where Arora’s remarks cut deeper than a single complaint about token bills. He is effectively saying that enterprise adoption will not be won by model mystique alone. Premium systems will need to show a concrete advantage on reliability, security, task completion or integration, otherwise buyers will treat them the way they treat any other over-specced software product.

Still, there will be pockets where higher prices survive. Complex coding, deeply integrated security workflows and regulated decision support are easier places to defend a premium than generic summarisation or first-draft content generation. Even there, buyers will increasingly ask for evidence rather than brand reassurance.

Anthropic, Google and OpenAI may still command higher prices for specialised coding, reasoning or regulated-workflow tasks. Even so, the burden of proof is shifting. Enterprises are asking whether the extra performance is material enough to justify the extra spend, particularly now that lower-cost and open-weight alternatives are getting close enough for many day-to-day jobs.

Arora’s intervention is not anti-AI. It is a market signal from a large enterprise software and cybersecurity buyer that the economics still do not work cleanly at scale. Buyers hear a warning about budget discipline. Vendors hear a demand to compress costs faster. Both sides are probably right.

That is why the next phase of the model war looks less like a race to the smartest demo and more like a race to the cheapest usable intelligence. Capability still matters. So does the bill.