The price of intelligence dropped again. Re-run your math.
As AI costs fall, smart product teams rethink architecture, features, and model choices - not just pricing.
JULY 13, 2026 • TEAM NFN
Few weeks ago Anthropic released Claude Sonnet 5. If you skim tech news, you saw the headlines: close to flagship-level quality, a one million token context window, and introductory pricing of two dollars per million input tokens through the end of August. The benchmark crowd spent the weekend arguing about coding scores. Fine. But if you are building a product right now, the benchmarks are not the story. The trendline is.
Every few months, intelligence one notch below the frontier gets dramatically cheaper. This has now happened enough times in a row that you should treat it as a planning assumption rather than a happy surprise. Here is what that means in practice.
First, re-run the math on the AI feature you shelved. If you priced out an AI feature a year ago and killed it because the per-user cost made no sense, that decision has an expiry date, and it may have already passed. We keep a list of features we have cut from client scopes purely on unit economics, and we revisit it every time prices move. Things migrate off that list constantly.
Second, read the fine print before you celebrate. Sonnet 5 ships with a new tokenizer that produces roughly 30 percent more tokens for the same text, which quietly claws back part of the price cut. The introductory pricing is set so the switch lands roughly cost-neutral for most workloads. The lesson is bigger than one model: never take a headline price at face value. Measure cost on your actual workload, with your actual prompts, before you re-plan anything.
Third, a one million token context window changes architecture decisions. A chunk of the retrieval plumbing that was mandatory two years ago is now optional for a version one. If your product's entire knowledge base fits in the context window, you might not need to build a vector pipeline before launch. Build it later, when scale demands it, with real usage data telling you what to optimize.
Fourth, and most important: do not weld your product to any single model. The right response to a market where prices fall and leaders rotate every quarter is to make the model a swappable part. At NFN Labs, every AI product we build treats the model layer like an engine you can lift out and replace. The workflow, the data, the evals, and the UX are yours. The model is a supplier.
There is an uncomfortable version of this advice too. If your product's entire value is a thin wrapper that calls a model and displays the output, the price collapse is coming for you as well, because your users can increasingly do the same call themselves. Falling model prices reward products where AI is load-bearing inside a workflow users cannot easily rebuild.
Cheaper intelligence does not favor whoever adopts it fastest. It favors whoever has the clearest judgement about what to do with it.
Sources: Anthropic announcement | TechCrunch | Simon Willison
