What Anthropic's AI spending reports mean for you

Anthropic has announced a set of reporting tools that show enterprise customers what their spending on its AI actually buys, according to TechCrunch. Managers can pick a date range and see what a team produced, what it cost, and where its workflows broke down. The view is built from internal analytics and employee surveys, and it comes with recommendations for improving results. That one view matters most for you. It turns a vague question, is this AI worth it?, into a report your manager can open.

The background you need

Tokens and token spend. AI models like Anthropic’s Claude read and write text in small chunks called tokens. A token is roughly a word or part of a word. Business customers are often billed by how many tokens their people use. Token spend is the company’s AI bill.

Think of it like an electricity meter. The meter tells you how much power you used. It does not tell you whether the lights were on in a room where someone was working, or in an empty one. The analogy breaks in one place. Electricity costs about the same no matter who uses it. With AI, two people can spend the same tokens and get very different results. One ships a finished feature. The other goes in circles. Until now, most companies saw the meter and not the result.

Agentic coding tools. These are AI tools that do more than answer a question. They take a task, such as fixing a bug, and work through the steps on their own: reading files, writing code, and proposing changes for a human to review. Anthropic’s version is Claude Code. OpenAI’s competing tool is Codex. These tools can run for a long time on one task, so they can use many tokens.

Return on investment (ROI). ROI means what you get back for what you spend. TechCrunch calls ROI a major concern among large customers. It also reports that most of the Fortune 500 are not seeing major gains from deploying AI, even as usage at venture-backed firms drives token spend.

Net revenue retention. This is a measure of how much more (or less) existing customers spend over time. TechCrunch reports that Anthropic told investors it had 900% net revenue retention. In TechCrunch’s words, spending from the same customers grew nine times between Q3 2025 and Q3 2026. TechCrunch also reports Anthropic recorded an operating profit in the second quarter.

Forward-deployed engineers (FDEs). These are a vendor’s own experts placed inside a customer’s company to help it get set up and get value from the product. According to TechCrunch, all the major AI labs are investing in these teams. TechCrunch describes the new reports as a kind of FDE-lite: a lighter, automated version of that hands-on help.

agent.md files. These are instruction files teams write so an AI coding agent knows their project’s rules and habits. Each tool handles them a little differently. That matters for switching costs, covered below.

What actually changed

Here is what is new, according to TechCrunch:

• Team reports. Managers can view a team’s work over a date range: output, cost, and where workflows broke down. The data comes from internal analytics and employee surveys.
• Recommendations. The reports suggest ways to improve results.
• Pooled group budgets. Teams can share a token budget, to manage spending better.
• A training default. Anthropic says it will not use this reporting data for model training by default.

Here is what was already true. Companies were already paying for AI by usage. Spending was already rising fast. Jellyfish, which makes tools for software teams, released a report based on 276,000 of its customers, as TechCrunch describes it. According to that report, spending on agentic coding tools has grown 10x since the start of the year. That spending now equals about 2% of total engineering cost. Claude Code and other tools built on Anthropic models were the most widely used. They were behind nearly a third of all proposed code changes submitted for review.

Jellyfish’s chief scientist, Nicholas Arcolano, told TechCrunch that “that 10x growth is growth that most CFOs didn’t plan for.” That is the pressure the new reports respond to. The bill has become big enough that someone has to justify it.

What it means for PMs, analysts and other knowledge workers

The reports change who answers the value question, and with what evidence. Before, a manager defending an AI budget had usage numbers and anecdotes. Now the vendor supplies a report that links cost to output and points to where work stalled. That has a few practical effects.

Your AI use may be judged on results. If your company turns these reports on, your team’s spending will sit next to what the team produced. Keep a short running record of what AI helped you finish. That way your own account matches, or corrects, what the dashboard shows.

Surveys are part of the data. The reports draw partly on employee surveys. What you say in them shapes how your team’s work is scored. Answer them specifically. Saved time drafting specs is less useful than naming which task, and roughly how often.

Breakdowns become visible. The reports show where workflows broke down. For a PM or analyst, that can help. It may show where an AI step keeps failing, so you can fix the handoff instead of blaming the tool.

Budgets may move to the team level. Pooled group budgets mean your team may share one pot of tokens. Expect conversations about who uses what, and about heavy users.

Switching gets harder. Jellyfish’s AI product lead, Nikolas Albarran, told TechCrunch that once a company has standardized its workflows on one tool, there is not a big enough reason to switch. He noted that each tool has its own quirks, such as how it handles agent.md files, and that “that cost to switch is not zero.” Reports that make one vendor’s tool easier to justify add to that pull. If you are part of a tool decision, count the setup effort, not just the price.

Albarran also told TechCrunch that OpenAI’s Codex has only recently started to show up in Jellyfish’s data, since OpenAI’s Astra model came out. According to TechCrunch, that release has so far not slowed Anthropic’s adoption numbers.

What we don’t know yet

• Who gets access. TechCrunch does not say which plans include the reports, or whether they cost extra.
• How output is measured. The report covers what a team produced. It is not clear how that is counted, especially for work that is not code, such as research or documents.
• How the surveys work. We do not know how often they run, what they ask, or whether answers are anonymous.
• The training default. Anthropic says it will not train on this data by default. It is not stated who can change that setting, or how.
• How reliable the recommendations are. There is no public evidence yet on whether following them improves results.
• Budget details. How pooled group budgets are set, shared, or capped is not described.