The $2 anchor that becomes a 50% trap

Anthropic just cut the price of its most agentic Sonnet model by 33%. The new Claude Sonnet 5 costs $2 per million input tokens and $10 per million output tokens. The old Sonnet 4.6 cost $3 and $15. That is a real discount, and it has a deadline.

The deadline is August 31, 2026. After that, Sonnet 5 jumps back to $3 and $15. That is a 50% increase from the intro rate. So you have roughly 60 days to build on the cheap number before the price goes up.

Here is the part most people miss. Sonnet 5 is now the default model for Free and Pro users. Millions of people will start using it without picking it. They get onboarded at the low anchor. Then the meter climbs. I think this is one of the smartest pricing plays I have seen this year, and I want to show you exactly why.

The Land-and-Expand Anchor

Every aggressive pricing move follows the same shape. Get in cheap. Get sticky. Then raise the number once leaving costs more than staying.

I call this the Land-and-Expand Anchor. You land the customer at a price so low it feels reckless to say no. Then you expand the switching cost until the higher price feels normal. The anchor is the first number a customer sees, and it sets what "expensive" means forever.

Anthropic set the anchor at $2/$10. That is the number developers will remember. When it moves to $3/$15 on September 1, most teams will already have agent workflows wired into Sonnet 5. Rebuilding on a rival costs engineering time, so they eat the 50% bump.

The trick is that the anchor is temporary but the habit is permanent. You feel the discount today. You feel the lock-in later. That gap between the two is where the whole strategy lives.

The Deep Dive: Why This Is a Sales Move, Not a Product Move

Let me put on the pricing hat. The question that matters is not "is Sonnet 5 good." The question is LTV:CAC. What does it cost Anthropic to acquire a developer, and what is that developer worth over their lifetime?

Sonnet 5 as the default for Free and Pro tiers is a customer acquisition machine. The acquisition cost drops toward zero because users get placed on the model automatically. No sales call. No demo. They just start typing. That is the cheapest acquisition channel that exists.

Now watch the lifetime value math. Anthropic's own enterprise figures show about $13 per developer per active day and $150 to $250 per developer per month, per the Finout and MorphLLM analysis. A single bug-fix task runs 25 calls and roughly $0.54 on Sonnet 4.6 at API rates. Agents burn tokens. That is the point.

Here is where the anchor earns its keep. The intro price makes the daily spend feel small. A full-day agent workload that costs around $594 per month on Sonnet 4.6 could drop toward the $400 range on Sonnet 5, assuming similar caching. So developers do not cut usage. They increase it. They run more agents because each one feels cheaper.

Then September 1 arrives. Price goes up 50%. But the usage habit is already baked in. Those agents are running in production, in Claude Code, in terminals, in browser tools. This is the golden goose logic. The goose is not the token. The goose is the workflow the developer built on your infrastructure.

I want to be honest about the weak spots, because a good offer survives its own damaging admissions. First, the discount is only 60 days. That is short. A rival could match $2/$10 tomorrow and the anchor loses its edge before switching costs pile up. Second, Sonnet 5 uses a new tokenizer, so the same text can bill 1.0 to 1.35 times as many tokens as Sonnet 4.6. That quietly eats part of the headline cut.

Third, one skeptic framed the whole launch as "cheap until August" tied to a possible IPO window. It is unclear whether this is a durable moat or a short capture play dressed up as one. The data is mixed. My read on this: the pricing is smart, but a discount is not a moat. The moat is the switching cost, and switching cost only forms if the workflows are genuinely hard to move.

But the capability numbers give the anchor teeth. On SWE-bench Verified, Sonnet 5 hits 72.7% versus Sonnet 4.6 at 62.3%, with Opus 4.8 at 79.4%. On Terminal-bench, tool use jumped from 55.4% to 76.1%, a 20.7-point gain. That is the biggest jump in the whole scorecard, and tool use is exactly what agents do all day. Near-Opus performance at less than half the Opus output price of $25 is a strong offer.

2031

Three signals inside the same shift

Pull back five years. The token was never the product. The default was the product.

The company that owns the default model owns the workflow. And the workflow is where compounding happens. Every agent a developer builds on Sonnet 5 is a small deposit into a switching-cost account that grows on its own.

We have watched this movie before. Costco sells the hot dog at $1.50 and makes it back on the membership. The cheap thing up front is not the business. It is the entry point to the business. Anthropic is running the same play with tokens as the hot dog and the workflow as the membership.

By 2031, I expect the agent infrastructure market to look like the cloud market did after AWS moved first. One or two defaults will hold most of the volume. The rest fight over scraps. Whoever owns the default in the land-grab window of 2026 owns an asymmetric advantage that competitors cannot cheaply reverse.

The risk cuts the other way too. If token costs keep falling across every provider, the anchor stops mattering. Models become interchangeable. In that world the moat is not price at all. It is the tools, the memory, and the agent SDK glued around the model. That is a different game, and it is unclear which one we end up playing.

What to Build This Weekend

Do not just read about the anchor. Test it with your own money and your own build. You have about 60 days before the price moves.

First, build one small agent on Sonnet 5 this weekend and log every token. An agent is just software that plans, calls tools, and acts on its own, not a chat window. Try Architect.new to spin up your first agent, since it has a Guided mode built for beginners. You do not need a CS degree for this.

Second, wrap the agent around a real task you repeat. Email triage. Ticket sorting. A bug fix. Use Blink.new or Lovable to stand up a simple front end so the agent has somewhere to live. Ship something ugly that works before something pretty that does not.

Third, measure two numbers. Cost per completed task at the $2/$10 rate. Then the same task recalculated at the September $3/$15 rate. If the 50% bump breaks your economics, you learn that now, not in a surprise invoice.

Fourth, rehearse the vendor conversation before you commit. Tough Tongue AI 2.0 lets you roleplay a hard talk, so practice defending why you would stay on Sonnet 5 after the price rises. If you cannot argue it out loud, you have no moat, just a discount.

Things will break. Agents loop. Tokens spike. One documented subagent fan-out produced a $47,000 bill, per Finout. Cap your spend, set a budget alert, and test aggressively. Get your reps in this weekend, and you will understand the anchor better than the people who only read about it.