Daily Issue #10 · 7 min read

Open models handle 1 in 5 AI tokens but earn 1 in 25 dollars. Revolut shows why that matters.

Menlo Ventures says frontier labs now take almost 70 percent of enterprise AI spending. Revolut's move to a self hosted open model shows what that number misses.

The short answer

Enterprises spent an estimated $176.8 billion putting AI to work in the year to September 2026, up from $33 billion in 2025, according to Menlo Ventures' annual State of Enterprise AI report, published November 3. Almost 70 percent of that money went to just three labs: Anthropic, OpenAI and Google. Open weight models now handle about one in five tokens but collect only about one in 25 dollars. On the same day, The Register reported how Revolut moved roughly 2,000 engineers onto a self hosted open model and says its AI coding bill fell to about a tenth of what it paid Anthropic and OpenAI. My view: both stories are true at once. Frontier models are winning the money because they win the hardest work. But a lot of daily AI work is not the hardest work, and the companies that notice this early will get far more out of every dollar.

What Menlo's numbers say

Menlo built its estimates from a survey of nearly 400 US enterprise decision makers plus a bottom up market model, and the full 2026 State of Enterprise AI report explains the method. Four findings stood out to me.

First, spending ran far past plans. The average company expected to spend 1.9 times what it spent last year and ended up at 4.7 times. The median company reached 2.2 times against a planned 1.8. Menlo puts the gap down to agents, which keep working in parallel after people step away, so cost follows the amount of work rather than headcount.

Second, the big labs are pulling away. Anthropic, OpenAI and Google hold 69 percent of deployment spending, up from 50 percent a year ago. Inside enterprise API spending, Anthropic grew from 40 to 55 percent and OpenAI from 27 to 31 percent, while Google fell from 21 to 9 percent.

Third, buyers are trading up. A year ago Menlo's most popular models included Claude Sonnet 4.5, at $3 per million input tokens and $15 per million output tokens. Today the most popular include Claude Opus 5.5 at $4 and $20, and OpenAI's flagship Astra model at $10 and $60.

Fourth, open models are busy but cheap. Their share of tokens rose from 11 to 19 percent while their share of spending slipped from 6 to 4 percent, because the average open model token fell 3.3 times in price to $0.13 per million. Closed model tokens average about $0.70.

Menlo concludes that capability, rather than price, is proving to be the deciding advantage. One thing worth knowing as you read that: the report marks Anthropic as a Menlo portfolio company. That doesn't make the data wrong, but it is a reason to read the interpretation with some care.

Why "1 in 25 dollars" undersells open models

Spending share is a strange way to judge a model that is cheap by design. If an open model does the same job for a fraction of the price, it will always look small in a chart of dollars. The token chart tells the more interesting story. Menlo estimates open model volume grew 22.5 times in a year, to about 2.9 quadrillion tokens a month, faster than closed model volume at 12.1 times.

This is not a story of open models taking over. Frontier models grew their token share too, from 18 to 42 percent. The group being squeezed is the smaller closed models, such as Claude Haiku and OpenAI's cheaper tiers, whose token share fell from 29 to 13 percent even after price cuts of 80 percent or more. So I read the data this way: the hard work is moving up to the frontier, and the cheap, high volume work is moving to open models. Only the first trend shows up clearly in a revenue chart.

What Revolut actually did

David Abitbol, who leads AI operations at Revolut, described the setup in a talk at AI Engineer NYC in October. The details are more useful than the headline.

It started with rationing. When Revolut rolled out Claude Code, Abitbol said some engineers spent "a thousand a month, 10,000 a month" without a clear return, so the company set strict budget limits. The limits fixed the bill but made people save their tokens for the hardest coding tasks.

Around April 2026 Revolut began self hosting Z.ai's GLM 5.1, and it later moved to GLM 5.3. It now serves about 2,000 engineers from 56 Nvidia B200 GPUs at a blended cost near $0.30 per million tokens. On Revolut's internal chart, coding time improved about 60 percent after the switch, and lead time, from picking up a feature to having it merged, improved about 70 percent. The conference write up notes that this comparison does not isolate the model change from other causes.

Claude Code and Codex are still available, but as the exception rather than the default. GLM is text only, so image tasks get routed to other models.

The line from the talk that stuck with me is about what cheap tokens change: "You no longer have to save your tokens for the end of the quarter or the most difficult task." Abitbol said the bigger gain was in lead time, not coding time, because engineers started using the model for the work around the code, like digging through the inbox and handling Jira and Confluence tickets.

Cheap tokens change how people work

Headlines framed this as Revolut cutting out Claude for a Chinese model. I think the more important point is about behavior. Per token billing pushes people to ration, and rationing kills exactly the boring, high volume uses that add up: summarizing a ticket, checking a log, drafting a test, answering a question about an old service. Those tasks rarely need the best model in the world. They need a good model that people feel free to call a hundred times a day.

That fits Menlo's data better than it first seems. The frontier wins where capability is the bottleneck, such as long agent runs and hard code changes. Open models win where volume and predictable cost are the bottleneck. A flat bill with no surprises at the end of the month is a feature in itself.

The fine print before you copy Revolut

I would not read this as "self hosting is always 10 times cheaper." Four caveats matter here.

• The $0.30 figure is Revolut's own. The conference write up says it reflects Revolut's deployment and utilization, not a portable price. A team with low or uneven traffic ends up paying for idle GPUs.
• Hardware alone wasn't enough. Abitbol said that just loading the weights onto GPUs gave poor performance. With the GPU count held at 56, tuning raised requests per second by 76 percent and cut time to first token from 14.26 seconds to 1.68 seconds. That took FP4 quantization, sizing the context window to real requests, cache reuse and separating the prefill and decode stages.
• Speed is what users feel. His target is a first token in under three seconds, and past five, he said, complaints start.
• You own the safety layer. Revolut relies on sandbox and approval rules that limit what a coding agent can run, plus detailed telemetry to investigate incidents. When you self host, all of that is your job.

A simple way to decide

If I were choosing a model setup for a company today, I would sort the work rather than pick one model. Keep the frontier for tasks where a better answer is worth a lot: complex refactors, long agent runs, anything customer facing and hard to undo. Move steady, high volume, text only work to an open model, either self hosted or on an inference provider, once the volume is big enough to justify it. Then measure lead time and real usage, not only the invoice. Revolut's numbers suggest the hidden cost of expensive tokens is the work people stop asking AI to do.

My takeaway

Menlo's report shows where the money is going. Revolut shows where a lot of the value might be hiding. The frontier labs will keep the hardest and most valuable work for a while, and they deserve to. But I don't read "1 in 5 tokens, 1 in 25 dollars" as open models losing. To me it shows a growing share of AI work being done cheaply enough that it barely shows up on the bill. If your team is rationing tokens, that is the number to think about.

Quick notes

• Stanford finds many famous LLM findings have expired: Stanford researchers Owen Queen and James Zou retested 50 widely cited claims about language models on 20 models released from July 2024 to September 2026. They class 44 percent as expired and 34 percent as weakened. Capability failures such as fabricated citations faded fastest, while flattery of users and judges favoring their own answers held up or grew stronger.
• OpenAI shows its models designing physical hardware: OpenAI said on November 3 that two team members used its Astra model to design a test device, including the enclosure, circuit board and software, in under a week. OpenAI says the device still needed several review rounds because the model did not consistently account for manufacturing limits and tolerances.

Common questions

How much did enterprises spend on AI in 2026?

Menlo Ventures estimates enterprises spent $176.8 billion deploying generative AI in the year to September 2026, up 5.4 times from $33 billion in 2025. The estimate combines a survey of nearly 400 US enterprise decision makers with a bottom up market model.

Which AI company has the largest share of enterprise API spending?

According to Menlo Ventures, Anthropic leads with 55 percent of enterprise LLM API spending, up from 40 percent. OpenAI has 31 percent and Google 9 percent. Menlo lists Anthropic as one of its portfolio companies.

How much of enterprise AI usage runs on open models?

Menlo estimates open models handle about 19 percent of tokens but only about 4 percent of model spending, because the average open model token costs about $0.13 per million against about $0.70 for closed models.

What model does Revolut use for AI coding?

Revolut self hosts Z.ai's GLM 5.3 as the default coding model for about 2,000 engineers, served on 56 Nvidia B200 GPUs. Claude Code and Codex remain available for exceptions, and image tasks are routed to other models.

Is self hosting an open model always cheaper than using Claude or GPT?

No. Revolut's blended cost of about $0.30 per million tokens depends on its own scale and utilization. Teams with low or uneven traffic, or without infrastructure skills to tune serving, may not see similar savings.