One-Week AI Cost Check: Reprice, Calendar, Test

A model price cut is an ops event

Your AI stack is still paying the old rate for some workloads. Google's Gemini 3.7 Flash launched on August 13, 2026. It launched at a price 50 percent below Gemini 3.6 Flash's launch price. The launch price is $0.75 per 1M input tokens and $3.75 per 1M output tokens. The introductory rate stays in effect until the year ends.

For a first dozen-person team, that window is useful because it makes the old spend visible. If your stack still routes work through the older model, you have three jobs: inventory the workloads on the old model, reprice them against the new rate card, and calendar the expiry with a migration test.

Token pricing is where the stack cost shows up. The rate card can change while the workload stays the same, so the discount needs a visible line item in the ops calendar.

The inventory decides what can migrate

Start with the work, not the vendor page. A model discount only matters if you know which jobs are paying the old rate. In a first dozen-person company, that usually means a handful of internal tools, a support summarizer, a coding assistant, or a data extraction flow.

The inventory should be boring. It separates workloads that are ready for migration from workloads that are still hidden in a shared account. Each row needs one owner, the old model, the endpoint, the monthly spend, and the new-rate estimate.

The model accepts a 1,048,576 token context window and can return up to 65,536 completion tokens. That matters for long-document jobs, but it does not remove the need to test.

The reprice step turns the discount into a number

If a workload is mostly input-heavy, the savings may be large. If it is output-heavy, the math changes. The reprice step makes that difference visible before you move production traffic.

Do not assume the old model disappears. Gemini 3.7 Flash remains fully supported in Google's Gemini API. That support gives you room to run both models side by side for a short period.

If you need a quality signal, note that Gemini 3.7 Flash scored 43.6 percent on FrontierCode, a coding benchmark, while the prior model scored 34.4 percent. That is not a replacement for your own test, but it tells you the new model is not a cheap downgrade by default.

The expiry needs a calendar date and a test

The price is the easy part. The expiry is the risk. The team needs a visible date before the discount disappears.

The migration test should be small enough to survive a busy week. It catches prompt breakage, latency drift, and cost surprises before the rate changes.

The test does not need to be perfect. It needs to be fast enough to finish before the team forgets why the discount mattered.

The checklist below is the minimum set. Each item is a check you can run in a minute, not a project.

• Check that every workload on the old model has one owner and one row in the inventory.
• Verify each row shows the old model, endpoint, monthly spend, and new-rate estimate.
• Flag the biggest savings for migration.
• Put the discount expiry in the team calendar with reminders.
• Schedule a short migration test for one production workload before the expiry.
• Write a go/no-go note for moving each workload.

Keep the checklist next to the ops calendar. If the test has not run before the year-end rate changes, the next meeting should be about the gap, not the model.