Your AI invoice: how model choices drive cost.

Sundar Pichai's most useful line at Google I/O yesterday wasn't about a product. It was about your AI invoice.

Something I hear myself saying a lot is that you have to pick the right model for what you are trying to do. There is no one-size-fits-all, and running everything at the best frontier model at max reasoning isn't the solution.

Sundar told the room that many companies are already blowing through their annual AI budgets, and we're only in May. His pitch: shift 80% of your workloads to Gemini 3.5 Flash and you could save over a billion a year. Worth noting Flash 3.5 itself is roughly 3x the price of its predecessor, so even Google's 'cheap' option is getting more expensive.

The Financial Reality

Set aside the vendor angle for a moment. The underlying claim is real.

• Uber's CTO confirmed in April that the company burned through its entire 2026 AI budget in four months.
• Heavy users on Claude Code were costing between $500 and $2,000 a month each.
• Adoption inside the 5,000-engineer organisation jumped from 32% to 84% in three months. Productivity went up. The bill went vertical.

That's not an Uber problem. It's an industry pattern. Meta experienced a massive, unmonitored spending blowout due to an internal gamified leaderboard called "Claudeonomics". Employees competing to climb the leaderboard consumed a staggering 60.2 trillion tokens in a single 30-day window. At retail API rates, on Opus, this would have cost ~$900 million.

Agentic tools consume tokens at an order-of-magnitude faster rate than chat, and finance teams budgeted on the chat curve.

A Matter of Engineering Discipline

Almost nobody needs frontier reasoning for the bulk of their work. Code completion, summarisation, classification, routine drafting, internal lookups: all of it runs perfectly well on cheaper, faster models.

Reserving the expensive ones for the jobs that actually need them is a basic engineering discipline applied to a new input.

The companies that work this out first won't be the ones with the biggest AI budgets. They'll be the ones treating model selection as an active decision rather than a default setting.

Pichai's pitch was Flash. The real question is wider. What's your routing logic?