The AI labs have a unit metric. Most engineering orgs still don’t.

In the same week, two very different conversations asked the same question. Wall Street asked whether the biggest AI lab’s growth is worth what it costs. CEOs asked whether their own AI spend is producing anything. In the super intelligence era, both come down to unit economics, and to measuring AI ROI with data instead of anecdotes.

What Anthropic told investors

At a closed investor meeting in San Francisco on October 14, ahead of a planned IPO as soon as November, Anthropic shared its third-quarter numbers, according to reports from the Financial Times and Business Insider. Revenue reached $16.8 billion, up about 45% from $11.6 billion in Q2, which itself had jumped from $4.7 billion in Q1. The company reported its second straight quarter of adjusted profitability.

By almost any standard those are remarkable numbers. Yet the reaction was mixed. The slowdown from triple-digit to 45% quarterly growth disappointed some prospective investors, and several reportedly expect bankers to lower valuation expectations, possibly below $1.5 trillion. The reporting also highlighted the pressures underneath: AI developers are cutting prices to win customers, Anthropic outlined more than $500 billion in compute and infrastructure spending over the next decade, and it expects revenue from rival labs distilling Claude outputs, estimated at 5% to 7% of revenue, to fall now that it is curbing the practice. AI hardware stocks extended a multi-day selloff in the days around the news.

The most telling detail was a unit metric. Anthropic reportedly generated $1.70 in revenue for every dollar it spent on compute during the quarter. Investors no longer want to hear only how big AI is. They want to know how efficiently it turns cost into value.

Every CEO is asking the same question

The same week, GitLab CEO Bill Staples put the enterprise version of this question plainly on LinkedIn: companies are spending more on AI, so where is the return? His answer is that tokens consumed, code generated and developer activity do not tell you. In software engineering, he argues, the unit of value is a change that makes it safely into production.

That leads him to a metric he believes will be one of the most important of this era: cost per accepted change. How much does it really cost to move a change from intent through implementation, review, testing, security and acceptance, and is AI bringing that number down? He shared that GitLab has introduced early access to Duo Agent Platform Impact Analytics at Transcend, connecting AI consumption to workflow results by team, task and model.

What cost per accepted change really includes

The idea is simple. The hard part is being honest about both halves of the fraction. Cost is not just the AI bill, and an accepted change is not just a merged pull request.

Get the definition wrong and the metric flatters you. If you count only token spend, AI always looks cheap. (Our comparison of AI coding ROI platforms shows how differently vendors define it.) If you count every merged PR as accepted, changes that get reverted two weeks later still look like wins. The honest version counts all the cost that a change actually consumed on its way to production, and only the changes that stayed there.

Cheaper tokens do not mean cheaper changes

The price war between AI labs is real, and it is one of the pressures investors flagged around Anthropic’s results. It is tempting to assume that falling prices will lower your cost per change automatically. They will not.

Context is the multiplier. Waydev Brain keeps all of it inside your walls.

There is one lever that moves every part of the cost per accepted change at once: context. An agent that knows who owns a service, which patterns your team standardized on, what broke last quarter and why, writes changes that pass review the first time. An agent working blind writes changes that look plausible and come back. Same model, same token price, very different unit cost.

The value of that context does not grow in a straight line. Each layer makes every other layer more useful. Ownership data is good. Ownership data joined with review history, incident history and delivery metrics lets an agent reason about risk, not just syntax. That is why the return on AI compounds as context gets richer, and why most organizations never see it: their best context is the context they are least willing to send to a third-party API.

That is the problem Waydev Brain is built to solve. Waydev Brain is a private, air-gapped deployment of Waydev for enterprises: its own model, Waydev 15B, built for the software development lifecycle, its own agents, and your engineering data, all running inside your environment. Nothing leaves your perimeter, so you can give it the context you would never paste into a public model.

Brain also fits the agents you already use. Through Waydev MCP, coding agents can pull engineering intelligence context such as AI impact, ROI, token usage and checkpoint results, which improves accuracy and reduces cost. Inside Brain, that same context is served from within your perimeter.

Building your cost per accepted change

You do not need a perfect number on day one. You need a consistent one you can trend. These are the inputs we recommend engineering leaders put in place.

From hypergrowth to unit economics

Anthropic’s quarter is a signal that the AI market is moving from a story about scale to a story about efficiency. The same shift is coming to every engineering budget. Leaders who can show their cost per accepted change falling as AI use grows will keep their budgets and earn more. Those who can only show token counts and lines of code will be asked harder questions.

Staples ends his post with a line worth repeating in your next planning meeting: the winners in AI will be the ones that turn intelligence into trusted outcomes most efficiently. That applies to the labs. It applies to your engineering org just as much.