Claude Code Isn't the Product. It's the Top of the Funnel.

You pay $200 a month for Claude Max and burn through what would cost $2,000 or more at API list price if you paid per token. It feels like a steal. It’s supposed to feel like a steal.

Here’s the question worth sitting with: if Anthropic genuinely didn’t care how much compute you burned in Claude Code, why did they spend the first half of 2026 fighting, in public, over exactly that?

The gift with strings you didn’t notice

The story developers tell themselves goes like this: Anthropic doesn’t care how much compute you burn in their first-party app. They’re subsidizing the workflow because the workflow is marketing. Claude Code teaches you their tool-calling patterns, their agent loop, their conventions, and then you go build a real product on the Anthropic API, where they make their actual money.

That story is mostly right. But “mostly right” is doing some work, and the details matter more than the slogan.

The numbers that make this more than a vibe

Anthropic’s total annualized revenue went from about $1 billion in December 2024 to roughly $47 billion by May 2026. That’s not linear growth, that’s a company being rebuilt in real time: $9 billion by the end of 2025, $14 billion by February 2026, then more than tripling again in three months.

Multiple 2026 analyses (Value Add VC’s breakdown of Anthropic’s business model is the most cited) put pay-per-token API revenue at somewhere around 75 to 85 percent of that total. Consumer subscriptions, which is where most Claude Code billing actually lives, are estimated at somewhere in the 5 to 15 percent range. As of late 2025 reporting, more than 300,000 business customers accounted for roughly 80 percent of revenue, and by April 2026 more than 1,000 customers were each spending over $1 million a year on Anthropic, a number that had doubled in under two months.

None of these figures come from an audited filing, because Anthropic is private. Treat them as directionally reliable, not gospel. But they’re consistent across enough independent sources that the shape of the story is hard to argue with: this is an API business with a subscription product bolted on the front, not the other way around.

Here’s the complication, though, and it’s a real one: Claude Code itself is not small. It reportedly hit around $2.5 billion in annualized revenue by February 2026, up from roughly $500 million just five months earlier. That’s bigger than the total revenue of most public SaaS companies. Calling something that makes $2.5 billion a year a “loss leader” undersells it. The honest framing isn’t “Claude Code makes no money, it’s pure marketing.” It’s “Claude Code is a real, large business, and it’s still a small fraction of an even larger, API-dominated whole, and it does double duty as the thing that gets you hooked on the patterns you’ll later pay full price for.”

The receipts: six months of Anthropic fighting over exactly this line

If the funnel logic were as simple as “we don’t care how you use compute,” there would be no reason to spend half a year adjusting the plumbing. Instead, watch what actually happened in 2026.

January 9, 2026. Around 2:20 AM UTC, Anthropic flipped a server-side switch with no announcement. Third-party harnesses that had been authenticating through Claude subscription OAuth tokens, tools like OpenCode, Cline, and RooCode, stopped working overnight. No warning, no changelog entry. Developers woke up to broken workflows.

January 15, 2026. George Hotz published a post arguing the crackdown was a strategic error. His point: choke the on-ramp too hard and people don’t come running back to Claude Code, they switch to a competing model provider that leaves the door open. It’s a direct challenge to the funnel logic. If the funnel gets pinched, users don’t queue up, they find another funnel.

February 20, 2026. The Register reported Anthropic “clarifying” the ban on third-party tool access, which is corporate-speak for confirming the January change was intentional.

April 4, 2026, 3 PM ET. Anthropic formally enforced that Pro, Max, and Team subscriptions no longer cover usage routed through third-party harnesses via OAuth. The specific loophole being closed: some tools were spoofing the claude-code-20250219 client header to impersonate the official CLI, letting them draw on subscription-priced tokens instead of paying API rates. Multiple outlets covering this framed it explicitly as closing a token arbitrage hole, not a security patch.

May 13 to 14, 2026. A partial reversal. Anthropic proposed an “Agent SDK credit” system: separate metered credit pools for programmatic and third-party use, priced at standard API rates, layered on top of the subscription (Pro getting $20 a month in credits, Max 5x getting $100, Max 20x getting $200), with a plan to fully separate that usage from the main subscription pool by June 15.

June 15, 2026. Anthropic cancelled its own plan. Agent SDK usage, claude -p, and third-party app access continued drawing from the existing subscription pools “for now,” with Anthropic saying it was reworking the approach to “better support how users build with Claude subscriptions.”

Crack down, half-fix, cancel the fix. Six months of visible thrashing over where exactly the line sits between “subsidized habit-forming usage” and “usage that should be paying full API price.” A company that truly didn’t care about this distinction would not spend six months adjusting it in public, walking it back, and then walking back the walkback.

Notably, actual Claude Code usage limits moved the other direction during the same window: on May 6, 2026, Anthropic doubled Claude Code’s 5-hour rate limits for Pro, Max, and Team plans and removed peak-hour throttling, citing new compute capacity coming online. Weekly caps, in place since August 2025, weren’t touched, and Anthropic extended a promotional 50 percent higher weekly limit through mid-July, a move one outlet described as a direct response to competitive pressure from OpenAI’s Codex.

Put those two threads together and the pattern gets specific. Claude Code itself, the actual first-party habituation product, got more generous limits. It’s the leakage around Claude Code, other tools riding on its subscription pricing without going through the product that’s supposed to build the habit, that got squeezed hard.

Steelmanning the skeptics

It’s worth taking the counterarguments seriously, because the funnel story is easy to overstate.

Hotz’s critique still stands as a real risk: aggressive enforcement can push developers toward providers that don’t bother enforcing this boundary, which would undercut the whole strategy. Whether that happened at scale by mid-2026 isn’t something the public data settles either way.

On a developer forum thread that circulated widely in 2026, one commenter called Claude Code “not a profitable business, it’s a loss leader,” estimating something like $8 a day in subscription cost against $100 a day of equivalent API usage, a roughly twelve-times subsidy at the extreme end. Another commenter pushed back hard: that comparison mixes a flat subscription against metered API pricing, and if there’s a subsidy happening, OpenAI’s reported infrastructure losses suggest this is an industry-wide feature of how inference gets priced right now, not something unique to Anthropic’s strategy.

There’s also a more optimistic read from AI-industry analysis outlet SemiAnalysis, which argued Claude Code matters less as a developer-acquisition funnel and more as proof that agent orchestration, not raw token consumption, is where AI value actually shows up, pointing to metrics like Claude Code’s share of GitHub commits (reportedly around 4 percent as of mid-2026, with projections toward 20 percent by year end) as evidence of genuine workflow displacement, not marketing spend.

And here’s the honesty check worth stating directly: no Anthropic executive has gone on record calling Claude Code “top of funnel” or an acquisition tool. That framing, the one in this post’s own headline, is an outside inference built from pricing behavior and enforcement patterns, not a confirmed strategy. It’s a reasonable read of the evidence. It is not something Anthropic has admitted to.

What this actually means if you’re building on it

If you’re a developer using Claude Code today with an eye toward eventually building something that calls the Anthropic API directly, you’re exactly the customer this whole structure is built around, and you should expect the on-ramp to stay smooth. Anthropic has a clear, repeatedly demonstrated incentive to keep Claude Code itself generous.

If instead you’re routing other tools’ inference through your subscription’s OAuth token specifically to avoid paying API rates, expect that seam to keep getting watched. It’s the one place Anthropic has spent the most enforcement energy across 2026, and the June 15 reversal was explicitly framed as “for now,” not a permanent truce. Build assuming that boundary tightens again, not that it stays open.

The reframe

Claude Code isn’t the thing that makes Anthropic money in any straightforward sense. Even at $2.5 billion a year, it’s a fraction of a $47 billion, overwhelmingly API-driven business. What it is, concretely and by design or not, is the onboarding flow: the place where you learn the tool-calling patterns, build the muscle memory, and get comfortable enough with agentic coding that graduating to direct API usage feels like the obvious next step rather than a leap.

Anthropic spent six months in 2026 making sure that graduation happens through their API and not around it. That’s not a conspiracy theory. It’s just what their own behavior, dated and documented, says they were optimizing for.