On 4 August the Linux Foundation launched the Tokenomics Foundation, with about thirty founding members. The list mixes the companies buying tokens with the companies selling them: Accenture, Booking.com, BNY, Flexera, IBM, JPMorgan Chase, KPMG, Oracle, SAP and ServiceNow. Its stated purpose is to standardise how organisations measure and manage AI token costs, and how those costs connect to business value. The foundation’s board met for the first time on 30 July. A technical steering committee is still forming.
Nobody founds a standards body for a minor problem. J.R. Storment, the foundation’s executive director, named the problem when Fortune’s Sheryl Estrada asked him what made it necessary: “Credit pricing, seat licenses, and bundled ‘unlimited AI’ plans sever usage from cost, and you cannot allocate costs you cannot measure to understand value.”
He also said something that undercuts the assumption behind most 2026 budgets. “Per-token costs fell heavily during 2023–2025,” he said, but “they have leveled off—and new model token prices are rising.” He describes AI as now “the largest and fastest-growing line item on enterprise technology budgets.”
Two publications three days later showed why the measurement problem is harder than it sounds.
Configuration decides the size of the bill
On 7 August Databricks published its internal playbook for AI coding costs, written by Patrick Wendell and four colleagues, with review from infrastructure leaders at four companies: Stripe, Coinbase, Uber and Ramp. The cost curve, left alone, “will eventually overtake revenue,” they write.
Tuning their harness and cache settings cut the number of generated tokens by close to half, with no quality degradation they could observe. Routing requests to the cheapest model capable of handling them reduced average task cost by more than 30% while roughly matching the quality of the most expensive model in the set. And when a developer types something like “investigate and fix this bug”, that sentence is a negligible fraction of what actually reaches the model, because the bill is dominated by context the user never explicitly supplies.
Databricks also identifies a lever for improving the cost-efficiency of models. The company separates the intelligence frontier, where the labs compete on peak capability, from the efficiency frontier: the models with the best ratio of intelligence to price. The second, they argue, is moving faster. Their evidence is unusually concrete for a vendor blog. Stripe evaluated Opus 4.7, found no meaningful quality improvement over 4.6 at higher cost, and declined to make it available internally. Databricks measured cost regressions comparing Opus 5.0 to 4.8. Both companies built their own benchmarks because, in Databricks’ assessment, public benchmarks reflect real-world coding performance poorly.
Databricks sells the infrastructure this playbook requires — Unity AI Gateway and the Omnigent meta-harness — and describes its own savings figures as “directional, based on an informal survey of development teams.” The individual mechanisms are checkable; the aggregate percentages rest on the company’s own data.
Ten per cent of the staff, ninety per cent of the spend
The same day, the chief executive of Varick Agents published an account of what enterprise AI rollouts look like from inside the implementations his firm runs. The distribution repeats regardless of organisation size: roughly 5–10% become genuine power users, about 20% use the tools occasionally and inefficiently, and around 70% barely touch them.
One of his examples involves a company paying an eight-figure sum in dollars each year for tokens, where 10% of staff consume 90% of them. The arithmetic is straightforward: if the remaining 90% used AI the way the top decile does, spend rises roughly tenfold.
These figures come from a company that sells the remedy. They are anecdote from its own projects rather than survey data. The exact value therefore matters less than the structure: at that distribution, a per-seat average tells you little about cost.
The figure in your budget is not a forecast
Three things set an AI bill: what a token costs, who consumes it, and how the tooling is configured. The first is outside your control, the second is one most firms cannot yet report on, and the third can move the number by half.
The price is rising. Storment says the declines are over and new models cost more than the ones they replace. A forecast that assumes continued cheapening rests on an outdated 2023–2025 trend.
Most firms cannot yet report the distribution of consumption. If a tenth of the workforce accounts for ninety per cent of the bill, an averaged cost per employee is not a useful indicator. And the wider the rollout, the higher the bill.
Configuration can move the bill by half. Databricks retuned its harness and cache settings and cut generated tokens by close to half with no observable quality degradation. Techniques for limiting context growth are still being developed. Databricks builds AI systems for a living, so you would expect its own setup to be near optimal, and yet it took 50% out of the cost with relatively simple configuration changes.
An annual AI figure is therefore not a reliable forecast. The Tokenomics Foundation exists to make the first of those three things comparable across vendors.
Impact on the Polish market
Everything that makes up the playbook Databricks applied — a central list of permitted models, request routing, an internal benchmark run against your own code, control over harness configuration, cost measurement at the level of a single session — requires an in-house AI team. Databricks, Stripe, Coinbase, Uber and Ramp have one. Many Polish companies using AI through Microsoft licensing and a solution deployed by an integrator do not have those capabilities.
In those companies the harness configuration, the default model and the caching behaviour are usually set by the integrator, and the client receives an invoice. The integrator has no reason to hold that cost down: under consumption-based billing, more tokens mean more revenue.
The starting point is less about building your own AI gateway and more about obtaining the right data. Storment says that managing how contracts expose token-level consumption data “takes no new literacy” and shows what you are paying for. In practice these are questions for the supplier: which model is set as the default and how can we change it, how does caching work, and will we receive a detailed report of token consumption.
There is also an opportunity here for anyone selling into this market. The gap between what the frontier companies can do about AI cost and what a typical Polish enterprise can do about it is large and widening. That gap is a service line.
Before token costs enter the AI budget
A few questions. Answering them will produce a better estimate of AI cost.
- Unit. What exactly is being counted, and does your supplier count it the same way?
- Attribution. Can you split consumption by model, workload, team and project? Storment treats this as the starting point for a CFO.
- Distribution. What share of consumption comes from your top decile of users? What distribution are you planning for?
- Configuration. Who sets the default model, the caching and the harness configuration — you or your supplier?
- Direction. Does your forecast assume unit prices keep falling? Storment says they have levelled off and are rising in new models.
- Value link. Is the budget tied to an expected outcome, or is it a flat ceiling? The latter will not tell you whether the spend paid off.
Briefing
The EU AI Office began enforcing — it hired 38 staff and opened a whistleblower channel. The Commission expanded the office on 31 July and transparency obligations became enforceable on 2 August, carrying fines to €15M or 3% of turnover. So what: a complaint under Article 50 can now be filed by an employee rather than only a customer. Fortune
OpenAI says it cannot rule out Critical cyber capability in an unreleased model, and has paused work on it. On 7 August the company reported that internal tests of Astra showed advances in agentic coding and cybersecurity large enough that it could not exclude the Critical threshold in its Preparedness Framework. That threshold means the ability to write functional zero-day exploits in hardened production systems unaided, or to carry out an entire novel attack from a high-level goal alone. Every prior model, including GPT-5.6-Sol, sat at High. OpenAI has suspended the internal work on Astra that does not meet its strengthened security requirements. OpenAI
Google rated a Gemini CLI flaw at the top of the CVSS scale. Novee Security showed that a public GitHub issue, filed by someone with no repository privileges, could reach CI credentials across coding agents from all three major vendors. The demonstration ran on each vendor’s default workflow configuration. The Hacker News
Summary
On 4 August thirty companies, including JPMorgan Chase, IBM and Accenture, founded an organisation to determine how token costs are measured and attributed. Its executive director says per-token prices have stopped falling and are rising in new models. Databricks, together with Stripe, Coinbase, Uber and Ramp, showed the same work costing half as much after a change of settings. Enterprise rollouts show a tenth of the workforce consuming ninety per cent of the tokens. Three things set an AI bill: the price of a token, who consumes it, and how the tooling is configured. You do not control the price, you usually do not have consumption broken down by person, and configuration can move the bill by half. The annual figure in a budget is therefore not a reliable forecast but a ceiling you might reach in March, or in November. In Polish conditions the strongest cost lever sits with the integrator who configures the tooling. Start by writing access to consumption data into the contract.
Stay balanced, Krzysztof
Krzysztof Goworek is founder of Quintant — AI advisory that gets enterprises from experiment to production value.