Anthropic will give companies token reports: customer spending grew 9x in a year

Anthropic has launched a set of reports for enterprise customers. They show how much a company spends on tokens (the units of text you pay for when working with a model), what teams got for that money and where work got stuck. At the same time, news emerged that Anthropic itself is doing well. In the second quarter it reached operating profit, meaning its core business earned more than it spent. Net revenue retention, which measures how much money the same customers bring in a year later, came in at 900%. Customers who spent a notional dollar in the third quarter of 2025 are spending nine in the third quarter of 2026.

What managers will see

The reports are built on internal analytics and employee surveys. A manager picks a period and gets a picture of the team:

• what the team produced over those dates;
• what it cost in tokens and in money;
• at which steps workflows broke down;
• recommendations on how to get better results.

Anthropic promises that, by default, it will not use this data to train its models. The company has also added shared group token budgets. Previously, spending was scattered across individual people; now a team can allocate a single pool.

In essence, this answers the most common question from large customers: is the investment paying off? TechCrunch notes that token spending is growing mainly because of venture-backed companies. Most Fortune 500 companies do not yet see a noticeable return from AI.

A bill CFOs did not plan for

A fresh report from Jellyfish, which provides analytics for developers, fills out the picture. The company studied the use of agentic coding tools (those that write and edit code on their own) among 276 thousand of its customers.

• Claude Code and other tools built on Anthropic models remain the most popular. They account for almost a third of all proposed code changes submitted for review.
• Spending on such tools has grown 10x since the start of the year.
• It now makes up about 2% of all engineering costs.

Two percent is small, but a tenfold increase was not in the budget. Jellyfish chief scientist Nicholas Arcolano says that even if 2% looks like a small amount, most CFOs did not plan for growth like this. In his words, simply spending is no longer enough: companies need to optimize costs and get real returns from the tools. Anthropic's reports hit exactly that spot.

Why customers are not leaving for a competitor

This summer OpenAI released the Astra model, and it looked like it might push Anthropic aside. So far it has not. Nicholas Albarran, Jellyfish's head of AI products, says Codex (OpenAI's coding tool) was barely visible in their data for a long time and only began to grow when Astra came out. Judging by the social network X, the competition looks fiercer: everyone there says Codex is better. The real data tells a different story. Anthropic did strong work with enterprise customers, and the tool simply worked well. People put a lot of effort into building their workflows around it, and they have no sufficient reason to change everything.

For a large company, switching is harder than for a lone programmer. Albarran points to one more factor: all the major labs are investing in forward-deployed engineering (FDE). That is the name for teams that send AI consultants to work inside a client company and help with rollout. Anthropic's new reports can be seen as a lightweight version of that help, without people on site. According to Albarran, the longer a company works with one vendor, the harder it is to leave: every system has its quirks, for example in how it handles instruction files for agents. As a result, the cost of switching stays above zero.

What this means for those who pay

The agentic coding market is moving from a race over model quality to a fight for customer retention. The winner is whoever can explain to the customer where the money went. If you pay for Claude Code or similar tools, it is time to calculate token costs per team and per result, not just per person. The reports are new, and their usefulness has not been proven. But they make AI spending visible to the CFO, and that is a strong argument at contract renewal.