AI writes more of your code every quarter. Do you know what one feature costs you now?

Most engineering leaders can tell you how many AI seats they pay for and how many tokens they burned last month. Very few can answer the question the business actually asks: what does it cost us to ship a feature now, and is AI bringing that number down? That number is cost per feature, and it is about to become the most important metric in the AI SDLC.

This post starts with how to define and measure it. Then it explains why it matters now, using Tomasz Tunguz’s argument that AI labs like Anthropic are turning into clouds, and that AI spend is becoming a usage-based infrastructure bill like cloud compute before it.

Cost per feature, defined

Cost per task is the vendor’s unit. Cost per accepted change is the engineering team’s unit. Cost per feature is the unit the business understands, because a feature is what customers use and what the roadmap promises.

Three rules keep the metric honest. Count every change tied to the feature, including the fixes after launch. Segment by feature size or type, because a new billing flow and a settings toggle are not the same unit. And only count features that reached users and stayed, not features that were merged and then rolled back.

The cheapest task can make the most expensive feature

Frontier models can differ by as much as 3x in price per task (more on that below), which tempts every CFO to route all work to the cheapest model. But price per task says nothing about how many tasks a feature needs, how often the output is accepted, or how much human review and rework follows. Here is a simple hypothetical to show why.

In this example the cheaper model wins on the AI line and loses on the feature, because human time dominates the total. The real answer will differ by team, codebase and work type, and sometimes the cheaper model will be the right call. The point is that you cannot know without measuring the whole feature, not just the token invoice.

Five inputs for cost per feature

Why now: AI is becoming a cloud

In his newsletter “What if They’re All Clouds?”, Tunguz, a venture capitalist at Theory Ventures, puts Anthropic next to AWS, Azure and Google Cloud. In Q2 2026, Anthropic’s quarterly revenue of about $12 billion compared with roughly $42 billion for AWS, $39 billion for Microsoft’s Intelligent Cloud segment and $25 billion for GCP. It is the smallest of the four, but grew about 15x year over year, against 82% for GCP, the fastest of the hyperscalers.

His argument is that each cloud started with a wedge: Amazon with ecommerce, Google with search and ads, Microsoft with the operating system. Anthropic’s wedge is frontier models and the leading coding harness, Claude Code. It also has the biggest infrastructure bet of the group: about $518 billion in cloud, compute and infrastructure commitments over roughly a decade, nearly 8x its Q3 run rate, against roughly 1x to 2x for the hyperscalers, according to figures Reuters reported from the company’s prospectus and CNBC covered.

Tunguz applies that lens to valuation. Using Wall Street estimates, including Morgan Stanley’s view of Google Cloud’s growing share of Alphabet profits as reported by TheStreet and analysis such as this Yahoo Finance piece on Google Cloud, he estimates a standalone GCP at roughly $1 trillion to $2.3 trillion, which would put Anthropic at $1.5 trillion to $2 trillion on par with it. We will leave the valuation debate to investors. The part that matters for engineering leaders is the cost side.

Your coding agent is now a cloud bill

The adoption numbers show how fast this line item has become standard. According to Ramp’s spend data, 87% of businesses that buy a coding agent now buy Claude Code, up 56 percentage points in a year. Codex follows at 41% and Cursor at 35%, and many companies pay for more than one.

Like cloud compute before it, this spend is usage-based, and prices differ a lot between providers. Tunguz cites Artificial Analysis data showing the cost per task on its Intelligence Index: $5.98 for Claude Opus 5.5, $3.26 for OpenAI’s GPT-6 Astra and $1.99 for Google’s Gemini 4 Argon. That is the 3x spread in price per task behind the example above, and Anthropic is using its pricing power at the top of it.

This is exactly how the cloud era started. Bills grew faster than anyone could explain them, until companies built FinOps practices and began measuring cost per customer, per transaction and per request. AI in the SDLC is at that same moment. The question is no longer “how much are we spending on AI?” It is “what does each unit of delivered software cost us now?” That is why cost per feature, not the token invoice, is the number to put in front of your board.

Concentration risk, on both sides

Tunguz flags that Anthropic has the most concentrated business of the four: two customers made up nearly 25% of its revenue in 2025, as Reuters reported from the prospectus. Many engineering orgs have the mirror image of that risk. When nearly nine in ten coding-agent buyers use the same tool, a single vendor’s price change can move your cost per feature overnight. Teams that measure cost per feature across tools and models can respond with routing decisions. Teams that do not will simply pay the new price.

FinOps for the AI SDLC

If Tunguz is right that AI labs are becoming clouds, then AI spend will follow the path cloud spend took: fast growth, then scrutiny, then discipline. The organizations that came out of the cloud era strongest were the ones that tied every dollar of infrastructure to a unit of business value. The same will be true in the AI SDLC, and cost per feature is the place to start.