By its own rough estimate, Tenex spends between one and two dollars on AI tokens for every dollar it spends on its employees. Even so, Alex Lieberman of Tenex said the firm worries less about runaway AI bills than it did six to eight months ago. He then used falling model prices to question a bet some AI data companies are making on corporate customers.
Tenex does AI engineering work for large companies, the "Fortune 2000" in the episode's title. Lieberman and his Tenex colleague CJ Hess spoke with swyx, host of the Latent Space podcast, at AI Engineer New York, in an episode published October 21, 2026.
A yardstick for token spending
AI models read and write text in small chunks called tokens, and providers bill by the token. For a company that leans heavily on AI agents, those bills can become a major expense, and the conversation turned to measuring them against payroll.
The discussion sketched a rough scale. Most companies spend perhaps $10 to $200 per employee on tokens. At more AI-native companies, token spend is more like 10% of what they spend on people. For the average AI startup the two are about equal, and the heaviest users spend 5 to 10 times as much on tokens as on staff. The question then was where Tenex fit.
Where Tenex lands
There was no exact number to hand. Counting raw tokens, Tenex's estimate was somewhere between one to one and two to one, which drew a surprised "Really?" and the remark that it was "pretty high." On that scale, a services firm lands around or somewhat above the average AI startup, though short of the top tier.
Tenex now thinks a lot about when to reach for the biggest, most expensive model, and its costs have dropped sharply with Anthropic's Claude Opus 5.5. The discussion recalled an earlier stretch when everyone used Astra and Fable because they simply gave the best results. Astra is OpenAI's GPT-6 Astra; Fable is Anthropic's top-end model line.
Anthropic's launch announcement, dated September 22, 2026, priced Opus 5.5 at $4 per million input tokens and $20 per million output tokens. The company estimated that typical workloads cost about 40% less than on Opus 5, partly because of the lower prices and partly because the model uses fewer tokens per task. In one internal test described in the same announcement, rewriting the HAProxy software from C into Rust, both Opus 5.5 and Fable 5.1 passed nearly all existing regression tests, and Opus finished faster at 51% lower cost.
Earlier in October, a company in a different line of work described a similar arc. Thomas Sohmers, a co-founder of the chipmaker Positron, told The Cognitive Revolution that Positron's token spend had briefly overtaken its salaries, peaking above $100,000 a day after GPT-6 Astra arrived. He called Opus 5.5 the "saving grace" because, on many of Positron's tasks, it beat Astra at a quarter of the price. Sohmers drew a different lesson about model choice, though: he did not want anyone doing real development work on anything less than the best model. Tenex describes rationing the biggest one.
The token-maxing hangover
The conversation then turned to the "token maxing" phase that many companies had gone through since the participants last spoke, with leaderboards ranking who used the most AI. Those companies, it was noted, found they had run up too much cost, a reckoning described as entirely predictable. The question for Tenex was whether it had feared that outcome or had simply been getting a good return.
Lieberman said the firm was "for sure worried about it" as its headcount grew. Still, he said he did not feel that anyone on the team ever really bought into the "token-maxing zeitgeist."
The conversation admitted to a few personal splurges anyway, along with a defense of them: if the boss is paying, you might as well go nuts for a while, because the weird experiments teach lessons you keep after you scale back.
From cheaper tokens to data companies
Lieberman then offered a connection that he admitted "may sound crazy," promising it was related. Earlier that day, he said, he had been discussing the long-term prospects of the data companies that supply AI labs, naming micro1 and Handshake among them. These firms sell human expertise and data that labs use to train and test models. Handshake, for example, recruits people to review and edit model output, annotate datasets and assess performance in their fields. micro1 also licenses companies' operational records, such as support tickets and engineering documents, as training and evaluation material.
According to Lieberman, more than 90% of these companies' revenue comes from the labs. A number of their founders, he said, expect that to flip within a few years, with enterprises becoming the main buyers. For that to happen, he reasoned, enterprises would have to do a lot of fine-tuning, meaning further training of a model on their own data so it specializes in their work. And fine-tuning only makes sense if it delivers better quality or a lower price.
He found both hard to believe. Quality is a hard bet, he said, because "the bar is just going to keep moving on the frontier," so a custom model risks being overtaken by the next general release. Price looks hard too, because token costs on the leading models keep going down.
Those falling costs, Lieberman said, are also part of why Tenex is not concerned internally. He suggested the industry has hit a kind of "Pareto point" where costs keep coming down on a per-task basis. As a result, he argued, it is much harder today than six to eight months ago for someone to token-max and "really F the company" on cost. Six to eight months ago, in his view, was when you would have been really concerned.
The case for costs rising again
When Lieberman asked, "Do you agree with that?", the reply was a qualified yes. If future versions of Astra or Fable turned out to be very expensive but produced the best results, Tenex could see itself trending back toward heavier spending.
On balance, though, the argument was that competition in both open-source and closed models will keep pushing prices down. Meta's Muse Wonder, which the company announced on October 12, 2026 as a frontier model for AI agents, came up as a fresh example of a model that is at the frontier and "dirt cheap." Many economically valuable tasks, the reply went, can already be done with Opus 5.5. Would a model ten times smarter really justify its cost for "moving some cells in a spreadsheet"? Probably not, came the answer.