In the 2009 film Cloudy with a Chance of Meatballs (yes… 17 years already), a young inventor builds a machine that turns water into food and rains dinner on the small town where he lives. It works beautifully, and everyone is happy until the servings start to grow. Each order arrives bigger than the last, and the town cannot keep up.
Anthropic released a new model, Claude Sonnet 5, on June 30, 2026, and on paper it looks like a bargain: close to the quality of its more powerful mainstream cousin, Opus 4.8, for well under half the price. But there is a catch: Sonnet 5 uses a new tokeniser, the process that chops text into tokens, the billing unit of AI, and it splits the same text into about 30 per cent more of them than the previous version. The price per token has not changed, but the same task now produces more tokens, and so, in effect, costs more. You place the same order, in bigger servings, and pay more!
Another Claude spent the last couple of weeks in adverse weather. On June 30 the US government reversed course on the export controls it had imposed on Anthropic, and Fable 5 made its comeback on July 1, after 19 days offline. To soften the disruption, Anthropic is giving paying customers extra room to use Fable until July 7: for these few days Fable draws on only half of their weekly usage allowance, compensation for the time they lost (and I know many are frenetically taking advantage of that, at ungodly hours of the day… including yours truly, guilty as charged).
The rate card is only part of what AI costs a business. The rest is decided by how the words are counted, among other factors, and by whether a government lets the model run at all. That is the thread we follow in this week’s issue of The Compute.
Portion control
The price of AI can fall on paper while the total bill goes up in practice, and Claude Sonnet 5 is this week’s clearest example.
Start with the rate card. Sonnet 5 carries the same list price as the model it replaces, $3 per million input tokens and $15 per million output; the tier above it, Opus, costs $5 and $25 respectively. For the first two months Anthropic is charging a lower introductory rate of $2 and $10, which is what makes the new model look cheap. That rate ends on September 1, when it returns to the standard $3 and $15, a rise of 50 per cent.
Then there is the part the rate card does not show. Sonnet 5 uses a new tokeniser that turns the same prompt or output into about 30 per cent more tokens; the developer Simon Willison measured roughly 1.4 times as many for plain English and 1.28 times for Python code. The model also reasons before it answers by default, and that reasoning is billed too. Each token costs the same as before; the same job simply uses more of them.
The two effects meet in September. A workload processed in July looks cheaper, because the introductory rate offsets the extra tokens. Come September 1, though, the higher rate lands on top of the larger token count, and the same work can run 20 to 30 per cent above what it cost over the summer. Artificial Analysis, an independent testing firm, put a typical Sonnet 5 task at about twice the cost of the old Sonnet, and a little above Opus, with the price label still reading $3 and $15.
Companies are already feeling the wider version of this, and starting to work out what it means. Uber ran through its entire 2026 AI budget in four months, most of it on Claude Code, and now rations what its staff may spend. So the real question for anyone drawing up a second-half budget is which price to plan around: July’s, September’s, or whether to give up and trust the Jevons paradox, the old rule that things get cheaper with time, only for us to use much more of them…
Open season
When the price gap grows wide enough, customers start to leave, and some of the defectors are the ones a model maker can least afford to lose.
Open-weight models, the kind a company can download and run on its own machines rather than rent through an interface the way the big frontier models are sold, have caught up quickly; most of the cheapest now come from Chinese labs such as DeepSeek. Coinbase has begun shifting more work onto them, drawn by a cost gap that has become hard to ignore.
Lindy, a small San Francisco firm whose software runs AI agents all day, went all in. Its founder, Flo Crivello, migrated the entire company off Claude and onto DeepSeek and said the switch saves millions, the cost of running the models having grown larger than Lindy’s total payroll. It was neither quick nor clean: the firm spent six to nine months testing open models and described the migration as a hundred times more work than expected. Mr Crivello also said he would come back to Claude if Anthropic reduced its prices.
Even Anthropic’s biggest backer is watching the meter. Engineers at Amazon, which has put billions into the company, are reported to be building cheaper in-house copies of Anthropic’s models, trained to imitate them, to lower Amazon’s own bills before token-based pricing takes hold. Copying the teacher to save on the tuition, in effect.
The common thread is switching cost. Leaving a frontier model is possible, and increasingly worth it, but it is slow and it takes real engineering to hold quality steady. Raising prices (or at least resisting a cut) relies on this friction: a supplier can ask for more precisely because moving is such hard work. How long that lasts depends on how fast the open models keep closing the gap.
Tokenomics
The largest firms in AI have started investing with compute rather than cash.
They can do it because serving a model keeps getting cheaper at the margin. OpenAI has just cut its internal cost for answering guest users, the people who use ChatGPT without ever signing up, by more than half, the sort of saving that comes from better hardware and software and that, over time, has the potential to drag prices down across the market.
But OpenAI did something even more interesting back in May. Its chief executive, Sam Altman, offered every startup in Y Combinator’s current intake (about 169 companies) $2m of its tokens each in return for a slice of equity. On paper that is $338m of compute; in practice it costs OpenAI far less, because it is paying in a product whose marginal cost is a fraction of the retail price it is credited at, and still falling. The startup gets its largest early expense taken care of, and hands over an ownership stake for it. But the $2m is fixed in dollars, so as models grow hungrier for tokens, and Sonnet 5 is this week’s signal that they do, the same credit buys less real work and runs out sooner. When it does, the startup is paying cash, built on OpenAI’s rails, with nowhere cheap to go.
Nvidia is heading the same way from the hardware side: rather than trying to sell a young company a rack of silicon it could never afford, it is standing behind the smaller cloud providers that host these firms, underwriting their financing and taking a recurring cut of the cloud revenue that follows, on top of the original sale. A one-off sale of a chip becomes a share of everything the silicon goes on to produce and earn.
The trade is the same in both cases, and shrewder than it first looks. Each firm pays in the one thing whose cost falls every quarter, compute, and collects the one thing that does not, a claim on the customer’s future. It also settles where the next wave of companies is built, because a startup raised on OpenAI’s tokens or Nvidia’s chips does not lightly move to a rival later. Compute, in short, has become capital: these firms now invest their tokens the way a bank invests money, and a chip or a credit that once earned a single sale can earn a share of an entire company’s future. The cheapest input in AI is quietly becoming its hardest currency.
Wirelessly yours,
Ziad Matar
Co-founder, Tokenando
Editor-in-chief, The Compute
After Hours
This week has been mostly about football, and I tend to drift, as ever, towards the underdog; there were a few happy surprises (go Morocco) and a couple of small heartbreaks (au revoir Senegal). Away from the pitch, but still on the flat screen, we watched a documentary on the hacker Kevin Mitnick, a good reminder of how untamed and full of possibility the web felt in the 1990s, which is more or less how AI feels now. The soundtrack came from Fabrizio Cammarata, whom we enjoyed live at a friend’s birthday party a few weeks ago, and have caught up with on Spotify today.
References
Claude Sonnet 5, pricing and tokeniser // Anthropic (Claude Platform docs) / https://platform.claude.com/docs/en/about-claude/models/whats-new-sonnet-5
Sonnet 5 tokeniser, measured // Simon Willison / https://simonwillison.net/2026/Jun/30/claude-sonnet-5/
Sonnet 5 cost per task // Artificial Analysis, via Yellow.com / https://yellow.com/news/claude-sonnet-opus-token-costs
US lifts export controls on Fable 5 and Mythos 5 // CNBC / https://www.cnbc.com/2026/06/30/anthropic-says-trump-admin-has-lifted-export-controls-on-claude-fable-5-and-mythos-5.html
Uber’s 2026 AI budget // CNBC / https://www.cnbc.com/2026/06/26/openai-anthropic-new-ai-spending-reality-as-users-shift-to-efficiency.html
Coinbase moves to cheaper models // The Decoder / https://the-decoder.com/coinbase-joins-the-rush-to-chinese-ai-models-as-western-labs-face-a-pricing-stress-test/
Lindy switches to DeepSeek // The New Stack / https://thenewstack.io/lindy-deepseek-anthropic-switch/
Amazon distilling Anthropic models // The Decoder / https://the-decoder.com/amazon-engineers-are-reportedly-distilling-anthropic-models-to-cut-costs-before-new-token-based-pricing-kicks-in/
OpenAI cuts guest response costs // The Decoder / https://the-decoder.com/openai-reportedly-cut-response-costs-for-guest-chatgpt-users-by-more-than-half/
OpenAI tokens for equity, Y Combinator // TechCrunch / https://techcrunch.com/2026/05/20/sam-altman-makes-mic-drop-offer-to-every-y-combinator-startup/
Nvidia revenue-share and credit-support model // Nvidia (company blog) / https://blogs.nvidia.com/blog/nvidia-unlocks-ai-compute-at-scale-capital-partners-to-power-ai-infrastructure-buildout/
General Magic (history and documentary) // General Magic / https://www.generalmagicthemovie.com