The Model Got Half as Expensive. Your Bill Won't.

Google cut the price of its workhorse model in half yesterday and made it substantially smarter at the same time.

Gemini 3.7 Flash runs $0.75 per million input tokens and $3.75 per million output. That is exactly half of what 3.6 Flash cost. On agent and coding benchmarks it did not just hold the line, it jumped: DeepSWE from 48.6 to 65.3, AutomationBench from 17.0 to 30.4.

Now read the part in the fine print.

That pricing is introductory and runs through December 31. On January 1 it becomes $1.50 and $7.50.

The Business Case With a Detonator In It

Somebody at your company is going to build a model against those numbers this quarter.

They will take the new price, multiply it by projected volume, show a payback period that finally clears the hurdle, and get the project funded. The math will be correct on the day they present it.

Then the calendar turns and their unit economics double.

This is not a Google problem. Introductory pricing is normal, disclosed, and reasonable. It is a discipline problem on the buyer's side, and it is entirely avoidable with one habit: model the rack rate, treat the discount as upside.

If a project only works at promotional pricing, it does not work.

Stop Negotiating the Contract, Start Fixing the Workload

Here is the harder truth for anyone trying to manage AI spend.

Prices in this category are falling faster than you can renegotiate anything. Half, in one release cycle, with better benchmarks. Whatever you extract at the table will be smaller than what the market hands you by waiting a quarter, and the effort you spend extracting it is effort not spent on the part you actually control.

Databricks published its own numbers on that in a post by co-founder Patrick Wendell and four colleagues. Routing each request to the cheapest model capable of answering it cut average task cost by more than 30 percent while roughly matching the quality of the most expensive model. Tuning their agent scaffolding and caching settings cut generated tokens by almost 50 percent with no observed quality loss.

Nobody negotiated anything. They changed how the work runs.

Their framing is the one I would steal: chase the efficiency frontier, not the intelligence frontier. Most teams are paying premium rates to run a top model against tasks a cheaper one handles identically.

Your Bill Is an Engineering Artifact

If your AI spend is growing faster than your AI value, the instinct is to blame the vendor's pricing. Usually the cost is coming from your own choices.

You send too much context. Every request drags a payload assembled by somebody who never measured it. Compression and cache configuration are unglamorous work with immediate returns.

You use one model for everything. Classification, extraction, and summarization do not need the frontier. Routing is not exotic anymore.

Your tools are chatty. Verbose tool output gets fed back into the model and billed as input on every turn. Audit what your tools actually return.

Nobody owns the number. Spend with no owner grows. Databricks replaced hard budget caps with visibility and progressive friction, which is smarter, because a hard cap just moves the work to somebody's personal account.

What Falling Prices Actually Mean

Cheaper tokens do not mean a cheaper bill. They mean more usage, because things that were uneconomical last quarter clear the bar this quarter.

That is good. That is the whole point. The value of a 50 percent price cut is not that you spend less, it is that a workflow you shelved in March is now worth building.

But it only works if somebody is deciding which workflows those are. Otherwise falling prices just fund more of whatever you were already doing, at greater volume, with the same unexamined context payloads.

What I'd Do This Week

Model at rack rate. Build every business case at January pricing. If it only pencils at the promo rate, it is not a project, it is a pilot.

Route before you renegotiate. Find the three highest-volume tasks running on your most expensive model and test whether a cheaper one matches. That is a week of work for a permanent margin change.

Instrument tokens per task. Not monthly spend. Cost per unit of work. Monthly spend tells you the number got bigger. Cost per task tells you why.

Give the bill an owner with authority. Somebody who can change the routing, not just report the total.

Keep a list of shelved workflows. When the price drops again, and it will, you want a ranked queue ready instead of a scramble.

The Part That Matters

The model got half as expensive and twice as capable in one release.

Your bill will still go up, because the bill was never really about the price.

It is about what you run, how you route it, and whether anybody is watching. Those three were in your control the whole time.