On Monday, Anthropic published the launch page for Claude Sonnet 5. The model ships with benchmarks that place it within spitting distance of Opus 4.8 — the company’s own flagship — and a pricing card that reads, at first glance, like a concession to market pressure. Introductory pricing: $2 per million input tokens, $10 per million output tokens. Available through August 31, 2026. After which, the company noted in a single sentence buried mid-page, the standard pricing schedule applies.

That sentence is the only part of the announcement that matters. Everything else — the benchmark charts, the safety assessments, the glowing developer testimonials already accumulating on Hacker News — is a delivery vehicle for it.

The standard understanding of this release, if you skim the first forty comments on any forum, is that Anthropic is doing what every AI lab eventually must: slashing prices to compete. OpenAI dropped GPT-5 pricing last quarter. Google’s Gemini 2.5 Pro undercut both of them. The narrative is that AI inference is becoming a commodity, and commodity markets reward the buyer. What most readers miss is that the $2/$10 number isn’t a price cut. It’s a trial offer with a fuse.

The Window Is the Product

Anthropic has been precise about the timing. August 31 is 61 days from launch — just long enough for a procurement cycle to start, stall, and then panic. Enterprise buyers who move fast can lock in two months of capacity at roughly half of whatever the standard rate will be. Buyers who wait for benchmarks to settle, for internal evaluations to complete, for the inevitable third-party bake-offs to publish — those buyers will pay full freight.

This is not a discount for the disciplined. It’s a penalty for the prudent. And it works because the AI procurement officer in 2026 has been conditioned by two years of price declines to fear only one thing more than overpaying: missing the window. The psychology is a feature, not a bug.

A procurement manager at a mid-market fintech firm described the dynamic to me over Slack this morning: “We’ve got three models in evaluation and my CTO wants to wait for the comparison reports. But if the reports say Sonnet 5 is 90% of Opus at 50% of the price, and we didn’t buy during the window, I’m the one explaining the gap to the board.” He paused, then added: “I hate that this works on me.”

Half the Price Is Still a Number

The thing nobody seems willing to say out loud is that we don’t actually know the standard price. Anthropic hasn’t published it. The $2/$10 figure is, by definition, a temporary promotional rate. When a SaaS company runs a summer sale and doesn’t disclose the post-sale rate, we call that what it is: a marketing tactic that relies on the customer’s inability to price the alternative.

This isn’t unprecedented. AWS has run reserved-instance discounts for years; the difference is that AWS publishes the on-demand rate alongside the discount, so the buyer can calculate the spread. Anthropic has chosen not to. The company’s own pricing page for existing models — Sonnet 4.6 at $3/$15, Opus 4.8 at $15/$75 — suggests that the standard Sonnet 5 rate will land somewhere north of $3/$15 and south of $8/$40. But that’s a wide band, and a $5 difference per million tokens translates to real money at enterprise scale.

The opacity is itself a negotiating tactic. By the time the standard rate drops, the enterprise that signed up early will have spent two months integrating Sonnet 5 into production pipelines, fine-tuning prompts, and building internal tooling around its specific strengths. Switching costs will have accrued. The discounted price wasn’t a gift — it was a down payment on captivity.

The Market Job Nobody Applied For

There is a genuine argument for this strategy, and it’s worth giving it its due. Anthropic is a private company burning venture capital to stay competitive with two larger rivals. It needs revenue predictability. A promotional window that converts free-tier and pro-tier users into committed API customers generates exactly that: a cohort of paying users who onboarded fast and will be sticky. The market rewards the company for it. The stock of goodwill among developers is, for now, high enough to absorb the maneuver without complaint.

But the market is also, as it so often is, missing the second-order effect. If the industry standard becomes “introductory pricing with a hard cutoff and no disclosure of the terminal rate,” then every enterprise AI buyer is playing a game of musical chairs with their own budget. The winner is the one who guessed right on the discount window; the loser is the one who waited for due diligence. That is not a market that rewards accuracy. It’s a market that rewards credulity.

And it’s not clear why buyers are applauding it. If a car dealership advertised a 60-day price and refused to tell you what the car would cost on day 61, you’d walk. If a cloud provider did it, the FTC would receive a letter. But when an AI lab does it, the coverage frames it as a victory for cost-conscious developers. The asymmetry is the story.

None of this is to suggest Anthropic is acting in bad faith. The model is real, the benchmarks are impressive, and the $2/$10 rate is genuinely low for a model of this capability. But the offer is, structurally, a discount swap: cheap tokens now in exchange for a pricing decision deferred until the customer is too committed to walk. That’s not a price war. That’s a subscription trap dressed up in benchmark charts. And the people who should be most skeptical — the enterprise buyers with the most at stake — are, judging by the first day’s reactions, the ones applauding loudest.

Sources