On June 30, Anthropic released Claude Sonnet 5 — not its biggest model, its middle one. Model launches are usually noise for a business owner. This one is worth two minutes, because it quietly redid the math on what AI costs to run.
The numbers: Sonnet 5 costs $2 per million input tokens and $10 per million output tokens through August 31, then settles at $3 and $15. Anthropic's flagship, Opus 4.8, costs $5 and $25. So the mid-tier model is 60% cheaper during the intro window and 40% cheaper after — and early coverage from TechCrunch, Engadget, and The New Stack lands on the same verdict: its performance is close to the flagship's on the tasks that matter, especially the multi-step "agent" work that runs automations.
A million tokens, for scale, is roughly 750,000 words. For a few dollars, the model can read every email your business received this month and draft a reply to each one.
You don't buy tokens. Your tools do.
Fair objection: you're a business owner, not a developer. You will never see a token bill. But every AI tool you subscribe to — the chat assistant, the booking bot, the thing that writes your follow-ups — pays token prices behind the scenes and passes them on to you as subscription pricing, usage caps, and rationed features.
When the cost of near-flagship intelligence drops 40 to 60%, three things happen downstream. Vendors quietly upgrade the model under their product, so the same tool gets smarter without you doing anything. Features that were metered get more generous. And automations that didn't make financial sense for a vendor to offer small businesses suddenly do — which is why the cheap tier of your favorite tool keeps gaining abilities that used to be enterprise-only.
There's a more direct effect too: Sonnet 5 is now the default model on Claude's free and Pro plans. If you use Claude in the browser — the way most owners I write for do — the assistant you already have got a real upgrade last week, at the same price.
Every cycle, the mid-tier model catches last year's best — at a fraction of the price. If you priced an automation in 2025 and passed, you priced a different product.
What this means if you passed on AI last year
This is the part worth acting on. Somewhere in the last year or two, you probably evaluated an automation and said no. The AI phone answering felt too expensive per call. The draft replies weren't quite good enough to trust. The proposal-writing tool choked on anything complicated. Reasonable calls — at the time.
But those were verdicts on 2024 and 2025 models at 2024 and 2025 prices. The industry pattern has held for three years running: the "good enough" tier keeps absorbing what the expensive tier could do, then the price drops again. A no from eighteen months ago isn't a decision anymore. It's a stale data point.
The move this week
Pick one automation you dismissed — just one. The follow-up drafting, the review replies, the phone answering, the weekly report. Re-run the trial with whatever tool you tested before, or test the same job directly in Claude, which now runs the new model by default. Give it the same real-world task that failed last time and compare honestly.
If it fails again, fine — you've spent thirty minutes confirming your call, and you'll retest next cycle. But if it passes, you've found an automation whose only obstacle was timing. Cheap intelligence is not a prediction anymore. It's a line item, and it was just repriced.