Under the Pavillon de Breteuil, in the Paris suburb of Sèvres, there is a basement vault that needs three keys held by three different officials to open. Inside, under a set of nested glass bells, sits a cylinder of platinum and iridium roughly 39 millimetres tall, machined in 1878 before being adopted in 1889 by the first General Conference on Weights and Measures as the International Prototype of the Kilogram. It is popularly known as Le Grand K.
From 1889 until 2019 that cylinder was the definition of the kilogram, and every set of scales in every port, laboratory and pharmacy around the world was calibrated back to it through a chain of national copies.
Roughly once every 40 years, Le Grand K was brought out for comparison against six official copies. A century later, the prototype and the copies had drifted some 50 micrograms apart. While nobody could say which had moved, the cylinder weighed one kilogram because it was the International Prototype of the Kilogram, so it could not be wrong. If it had lost a few micrograms, then every object on earth was fractionally heavier when weighed against it.
In 2019 Le Grand K was retired, after the 26th General Conference on Weights and Measures defined the kilogram through the Planck constant, a fixed quantity in physics that does not drift. Any properly equipped laboratory can now derive a kilogram of its own from that constant, without sending anything back to Paris.
In my world today, the AI token has taken centre stage in many discussions, yet there is no vault, cylinder or conference to define it. An abstraction with a very precise meaning, a token is a fragment of text, roughly the size of a short word, that an AI model reads as it receives a prompt and generates as it builds an answer, counted in both directions and now used to meter a fast-growing share of what companies spend on technology. Every AI lab writes its own tokeniser, the software that defines its token dictionary, which means the same sentence is equivalent to a different number of tokens at OpenAI, at Google or at Anthropic. It even differs between models from the same lab.
The company that sells the tokens also decides what each one is, there is no Le Grand T to define that. This issue tries to make a bit of sense of the whole situation. I hope you enjoy it!
Bill drift
Anthropic published Claude Sonnet 5 on June 30, 2026 and noted in a footnote to the announcement that the model uses a new tokeniser, and that the same input now maps to roughly 1.0 to 1.35 times as many tokens as before, depending on the type of content.
The headline, though, was touting a price cut: Sonnet 5 was billed at $2 per million input tokens against the standard Sonnet rate of $3, which made it 33 per cent cheaper. Sending in a prompt or document at the top of that tokeniser range would mean 1.35 times as many tokens, so the same job could cost $2.70 instead of $3.00, and the real saving is 10 per cent rather than 33. Both numbers are true at once, however the one that applies depends on what the input contains, and the customer only finds out when the invoice arrives at the end of the month. The same footnote mentions that Anthropic had already made a similar change at Claude Opus 4.7, so the unit has been redefined twice within the span of one year.
A week ago, OpenAI released GPT-6 Astra on September 3, 2026 at $10 per million input tokens, but just $1 per million when the input is cached, and in the opposite direction, $50 per million tokens on output, or $12.50 per million to write to the cache. Batch and flex processing are priced at half those rates, while fast mode is priced at double. None of this is new: caching, batch tiers and priority tiers already exist across the major labs, Astra just inherited the pricing mechanisms, same as everyone else.
What Astra adds however is a cliff: above 272,000 input tokens the billing moves to a second price level, and the new rates apply to the whole input rather than to the part above the threshold. While a prompt of 272,000 tokens costs $2.72 to read, add just 1 token and a prompt of 272,001 tokens doubles in price to $5.44, that extra token alone cost as much as the 272,000 that came before it. Cliffs are not an OpenAI invention either: Anthropic used to charge double above 200,000 tokens, but flattened the price starting from Claude 4.6 onward, so Fable 5 and the rest of the current Claude line-up now price the full million-token window at standard rates.
But Astra also changes what the meter actually counts. The Information reported on September 2, 2026 that the frontier model uses recurrent depth (sometimes called a looped transformer), in which it passes information through its own internal layers repeatedly before producing any output at all. Reasoning models until now have been thinking out loud and charging for it, because the reasoning comes out as output tokens and adds to the bill at output rates. The recurrent depth thinking that happens inside Astra’s layers emits no tokens, so it does not impact the bill, which makes a better financial case for the companies buying AI, because it takes the reasoning cost out of their budget. Both Anthropic and Google DeepMind are now examining this technique.
Reverse tokenmaxxing
In early September 2026, Meta told its engineers that token consumption would no longer count towards their performance reviews. Under the previous system, introduced about a year earlier, engineers were graded on AI-driven impact and sorted into tiers labelled AI Native, AI First and AI Enabled. Engineers began burning tokens to lift their scores, and the practice acquired an internal name, which turned into industry jargon: tokenmaxxing.
The results are visible in Meta’s own figures. Code changes to internal systems rose 220 per cent year on year while changes that actually reached users rose only 36 per cent, and internal AI spending is heading towards billions of dollars in 2026.
OpenAI published its own consumption figures on September 6, 2026: 3.1 agent-workdays for every human employee day across the research organisation, with a median researcher spending more than $600 a day on inference at API prices, and the 90th percentile spending above $7,000 a day. Doing the maths over 250 working days the median researcher spends about $150,000 a year and the researcher at the top of the distribution about $1.75m, though OpenAI obviously does not pay the public prices, so we can assume this is an internal transfer priced at list.
Meta’s remedy is to point the targets in the opposite direction. The old count was part of each engineer’s review, where achieving a high number was the goal. From 2027 onwards, the company plans to implement usage budgets and a central cost dashboard, where tokens are counted against a cap, which must not be exceeded.
No outcome, no fee
The case against billing by the token is not that it is inaccurate, the issue is that a company putting in an AI agent to replace part of its support desk cannot be charged by the seat, because the seats are what it is removing, and it does not want to pay by the token either, which measures how hard the vendor’s model worked rather than what the company got out of it necessarily. Hence the emerging industry discussion about a move to billing by outcomes.
Salesforce, for instance, sells Agentforce, its platform for AI agents that handle customer service and sales work inside its own CRM, in three different ways.
The first is $2 per conversation.
The second is credit based: $500 buys 100,000 Flex Credits, and a standard action, meaning a record updated or a workflow run, costs 20 of the credits, so basically 10 cents. That puts 20 actions at the same $2 price as one conversation, and whether credits or conversations are cheaper depends entirely on how many actions a typical conversation turns out to need.
The third way is a flat licence starting at $125 per user per month, which takes us back to the traditional per-seat pricing model.
The credit deserves a moment of its own, because a token is at least something the model handled. A credit is a voucher the provider issues in its own “currency”, redeemable for work at a rate it publishes and can revise. Salesforce introduced Flex Credits in May 2025 and described them as tying spend to measurable, valuable AI-driven results, although what is being measured is the number of actions that its own software chose to perform.
Intercom publishes its price: $0.99 when its Fin agent resolves a customer conversation, and zero when the conversation goes to a person. Its documentation is unusually clear about how the count is made. A confirmed resolution is one where the customer says the answer helped. An assumed resolution is one where the customer stops replying for 24 hours. Both are billed at $0.99, so a satisfied customer and a customer who might have just given up produce exactly the same charge.
Zendesk on the other hand does not publish its rates. Buyers report around $1.50 per automated resolution on committed volume and about $2.00 pay as you go, and Zendesk counts a resolution only where the AI handles the matter end to end.
Outcome pricing was meant to take the provider-defined unit off the invoice, but so far it merely moves it from the tokeniser into the pricing page or contract negotiation process. The definition of a resolution, the length of the inactivity window and whether an escalation cancels the charge are all drafted by the vendor, and a large enough buyer can negotiate every one of them. While billing by outcome is a genuine improvement on a non-negotiable tokeniser, the price and the unit of measure seem to be even more abstract than the token they are trying to replace.
Le Grand K drifted 50 micrograms in a century, and replacing it took delegates from 60 countries, four days at Versailles in November 2018 and a unanimous vote, seven years after the conference had first resolved to find a better definition back in 2011. The token, the conversation, the resolution and the action are each defined by whoever sends the invoice, and no conference has yet been scheduled to align that.
A note to our readers
Tokenando is an official media partner of AI Summit Barcelona, and we will be covering it in The Compute. The second edition runs on September 22–23, 2026 at the World Trade Center Barcelona.
On Wednesday, September 23, 2026 at 16:15 CEST on the Panel Stage, I will be moderating Panel 19, “From BSC to Brussels: Europe’s Supercomputing Bet on AI”.
Readers of The Compute get 10 per cent off tickets with the code THECOMPUTE10 at aisummitbarcelona.com.
Wirelessly yours,
Ziad Matar
Co-founder, Tokenando
Editor-in-chief, The Compute
After Hours
In the past week we finished the latest season of Silo, binged both seasons of The Agency, and started The Bombing of Pan Am 103 and Jo Nesbø’s Detective Hole. I am not sure why I always drift towards dark series, but somehow I can’t avoid it.
The best of the week was off-screen: solving crosswords at 30,000 feet, with most of the cabin crew pitching in by the end. In other news, there was also a happy hour that started far too early, if that is even possible, and ran for six hours fuelled by ice cold Estrella Galicia and a few slices of pizza.
References
“Introducing Claude Sonnet 5”, Anthropic, June 30, 2026 (new tokeniser, 1.0 to 1.35 times mapping, Opus 4.7 precedent, $2 and $10 pricing made permanent August 10, 2026). https://www.anthropic.com/news/claude-sonnet-5
“GPT-6 Astra: A new generation of intelligence”, OpenAI, September 2026. https://openai.com/index/gpt-6-astra/
“OpenAI’s new reasoning technique alarms AI safety experts”, TechCrunch, September 2, 2026 (recurrent depth). https://techcrunch.com/2026/09/02/openais-new-reasoning-technique-alarms-ai-safety-experts/
“Researchers fear safety disaster ahead of OpenAI’s Astra release”, The Verge, September 2, 2026. https://www.theverge.com/ai-artificial-intelligence/988334/openai-astra-ai-monitoring-safety
“Meta Pushes Its New AI Agent on Employees, but Eases Off on Tokenmaxxing”, Wired, September 3, 2026. https://www.wired.com/story/meta-pushes-its-new-ai-agent-on-employees-but-eases-off-on-tokenmaxxing/
“OpenAI’s chief scientist says no lab should keep scaling at maximum speed”, The Next Web, September 6, 2026 (agent-workday ratio, daily inference spend per researcher). https://thenextweb.com/news/openai-slowdown-pachocki-alien-mind-research-intern-compute
“Agentforce Pricing”, Salesforce (conversations, Flex Credits, per-user licensing). https://www.salesforce.com/agentforce/pricing/
“Salesforce Introduces New Flexible Agentforce Pricing to Accelerate the Digital Labor Revolution”, Business Wire, May 15, 2025 (20 credits per action, $500 per 100,000 credits). https://www.businesswire.com/news/home/20250515332990/en
“AI customer service agent pricing comparison”, Intercom ($0.99 per outcome). https://www.intercom.com/learning-center/ai-customer-service-agent-pricing-comparison
“About automated resolutions for AI agents”, Zendesk (definition of an automated resolution). https://support.zendesk.com/hc/en-us/articles/5352026794010-About-automated-resolutions-for-AI-agents
“Historic Vote Ties Kilogram and Other Units to Natural Constants”, National Institute of Standards and Technology, November 16, 2018 (60 countries, unanimous vote at Versailles). https://www.nist.gov/news-events/news/2018/11/historic-vote-ties-kilogram-and-other-units-natural-constants
“Kilogram: Introduction”, National Institute of Standards and Technology (International Prototype of the Kilogram, triple-locked vault, 2019 redefinition). https://www.nist.gov/si-redefinition/kilogram-introduction
Insights in a bottle ! Efficient short and a joy to read !