On Monday, Anthropic shipped Claude Sonnet 5 and made it — not some watered-down Haiku variant, but the actual Sonnet model — the default for Free and Pro users. No gating behind the top tier. No waiting list. It went live June 30 at $2 per million input tokens and $10 per million output, undercutting its predecessor by roughly a third through August, after which the price snaps back to $3 and $15.

The benchmarks tell a tidy story: 63.2% on agentic coding, up from Sonnet 4.6’s 58.1%, closing ground on Opus 4.8’s 69.2%. The company is calling it “the most agentic Sonnet model yet.” The press coverage, predictably, is a consumer-welfare hymn: cheaper, better, more accessible.

And that is exactly the wrong thing to focus on.

The Free Tier Is a Data Flywheel Disguised as Generosity

What no one says out loud — what the pricing page doesn’t print but the positioning screams — is that Anthropic needs your behavior more than it needs your $20 a month.

Making a model this capable the default for free users is not a philanthropic gesture. It is a trade. You get frontier intelligence at zero marginal cost; Anthropic gets millions of new interaction traces — real-world agent behaviors, browsing patterns, computer-use sequences — at a scale no paid tier could generate. One engineer I spoke with on the floor of a San Francisco AI conference, nursing a coffee and staring at the announcement on his phone, put it bluntly: “They’re not selling me tokens. They’re buying training data, and I’m the vendor who doesn’t invoice.”

The numbers make the logic clear. At $2 per million input tokens, a free-tier user who burns through even a modest 50,000 tokens per session costs Anthropic roughly a dime. If that session produces a single BrowseComp-style agent trajectory that improves the next model by a fraction of a percent, the return dwarfs the cost. This is not a consumer product. It is a data-acquisition engine with a chat interface attached.

The Agent Business Is Not the Chatbot Business

Here is the thing the “cheaper AI for everyone” narrative politely ignores: Anthropic is not optimizing for you to have nice conversations. It is optimizing for agentic behavior — the kind where Claude books your flights, fills out your expense reports, navigates your browser, and interacts with other software on your behalf.

That is a fundamentally different product from a chatbot, and it carries fundamentally different economics. A chatbot monetizes through subscriptions. An agent platform monetizes through integration — through being the middleware that sits between a user and every service they touch. The free model is a beachhead. Once you’ve trained your workflows into Claude’s agent loop, switching costs rise sharply.

Look at the pricing structure again. The promotional rate expires August 31, at which point Sonnet 5 reverts to exactly Sonnet 4.6’s pricing — $3 input, $15 output. But the tokenizer burns roughly 30% more tokens for the same text, according to NextFuture’s early analysis. So the “same” price is actually a 30% hike on real-world workloads come September, at which point the free tier has hooked enough users that the paid tier absorbs the increase quietly. The discount is not a discount. It is a trial period for a more expensive product.

What the Benchmarks Actually Tell Us

Five percentage points of improvement on agentic coding benchmarks is real but incremental. The gap between Sonnet 5 (63.2%) and Opus 4.8 (69.2%) is still six points wide. That is a meaningful delta, and it tells you something about how Anthropic views its own model hierarchy. Sonnet is not approaching Opus parity so much as it is being positioned as “good enough to deploy at scale.” Opus remains the premium product for people who need the last six points badly enough to pay for them. Sonnet is the volume play.

The volume play makes sense if and only if the long game is platform lock-in. A startup CTO I traded DMs with on Tuesday described it this way: “I can build on Sonnet 5 for free, prove the thing works, and then when it’s mission-critical, I’m paying whatever they ask because rewriting the agent stack for another model would cost more than the price hike.” That is not a customer relationship. That is a moat.

None of this makes Anthropic evil. It makes them rational. The company is executing exactly the playbook that any smart platform business would execute in a land-grab phase. The mistake is treating the pricing as a consumer-friendly move rather than what it is: a structural shift in how AI companies capture value. Free users are not the customer. They are the raw material.

If that sounds too cynical, ask yourself why the most agentic Sonnet ever built — the one that can browse the web, operate your computer, and chain together multi-step tasks — launched first to the tier that pays nothing.

Sources