OpenAI recently disclosed how much its own researchers spend running AI models against their work. The median researcher now spends around $600 a day on tokens, roughly $12,000 a month. The top 10% spend more than $7,000 a day, close to a staggering $140,000 a month.
Cathie Wood, founder of ARK Invest, posted about the figures on 8 September, calling them evidence of a “massive productivity” shift and pointing to researchers writing more code, running more experiments, and deploying several AI agents at once.
It’s a striking number, and it raises a real question for any UK business watching AI spending climb: is that kind of usage a sign of productivity, or just a sign of wanton spending?
The number is real, the conclusion isn’t confirmed
Worth being clear about what OpenAI actually published.
The token bills are real, denominated at standard API rates. What’s less settled is the productivity claim sitting on top of them – OpenAI defined its own measure of research output, ran its own comparison, and published its own conclusion. No outside lab or academic auditor has verified the ratio between token spend and research progress. This doesn’t make the spending data useless, it just means the “massive productivity” framing is OpenAI’s interpretation of its own numbers, not an independently confirmed finding.
This distinction matters more for UK businesses than the headline figure does. Token consumption tells you how much a team is spending on AI but it doesn’t automatically tell you whether that spend produced anything a business would pay for on its own terms.
What actually counts as evidence
To separate genuine productivity gains from higher activity, a business needs an output measure that exists independently of the AI tool: work delivered, defects avoided, time saved on a task with a known baseline.
Lines of code written or experiments run don’t qualify on their own, since AI agents can generate both in volume without necessarily producing more finished, validated work. The useful comparison is cost per validated outcome against what the realistic alternative would have cost in salary, time, and missed
opportunity, not cost per token against some vague sense of busyness.
We’re leading the way on innovation.
Dr Logic has performed transformational work for clients across the UK, unlocking new levels of efficiency and security for multiple industries.
UK data backs up why this distinction matters. The Department for Science, Innovation and Technology found that 75% of UK businesses using AI report a productivity gain, but only 12% report an increase in revenue. That gap between felt productivity and measured financial return is what some researchers are now calling the UK’s “productivity-profit gap,” and it’s exactly the pattern a business should expect to see if token spend and genuine output aren’t yet the same thing.
The agent economics question
If one skilled person can run several AI agents in parallel, effectively supervising a small team of automated workers, that does change the maths of knowledge work. Fewer people could plausibly produce more output, at least in roles where the work breaks down into tasks an agent can execute and a human can check. But so far, that shift hasn’t shown up in UK employment data.
The Office for National Statistics found that only 4% of UK businesses using AI report a workforce reduction as a result, and the British Chambers of Commerce puts the figure at a similar level: the overwhelming majority of adopters say AI is supporting existing staff, not replacing them. Whether that holds as agent capability improves is an open question. It hasn’t happened yet.
Calculating your own ROI
The framing that $140,000 a month could be cheap, if it replaces months of work otherwise done by a team, is the right question for any UK business to ask about its own AI spend, at whatever scale.
The answer depends on three things:
- What the AI actually finished (not attempted)
- What that would have cost through the normal route
- Whether the people using the tool know how to specify tasks precisely enough to get validated output rather than plausible-looking output.
OpenAI’s own top-spending researchers understand model behaviour well enough to catch subtle errors most people wouldn’t notice. That skill, not the size of the token bill, is what makes their spend defensible. Most businesses adopting AI tools don’t have that skill yet, which is also visible in the UK numbers: 46% of small firms told the Federation of Small Businesses they simply lack the knowledge to use AI effectively.
How transferable is any of this?
Elite researchers at a frontier AI lab are not a representative sample of the UK economy, and the gap shows in the adoption figures. ONS puts general AI use across UK businesses at 23%, the BCC’s broader definition puts it at 54%, but strategic deployment with a defined business purpose sits at around 16%. The UK government’s own assessment is that a full, well-managed embrace of AI could add roughly £47 billion to the economy over a decade through productivity gains. That’s a meaningful number, but it depends on the “well-managed” part doing a lot of work, and right now most SMEs are using AI for lighter tasks rather than the kind of agentic, multi-step work OpenAI’s researchers are running.
If you want help figuring out where your business actually sits on that gap, whether your AI spend is buying real output or just more activity, get in touch and we’ll take a proper look with you.
