Google's cheap new model is built for agents. I gave it four days of my actual work.

On Friday I pointed Google's newest model at a small, real, slightly embarrassing job: a side project of mine that has a failing test I have been ignoring for a month. Not a demo repo. My repo, with my bad variable names and a TODO from March that just says "fix later, coward." I gave Gemini 3.7 Flash the failing test, the error, and a one-line instruction — make it pass without breaking the other twelve — and let it work. It fixed the test in about ninety seconds. It also, in the same pass, quietly rewrote a helper function I had not asked it to touch. Both of those things are the story of this model, and the gap between them is what you are actually paying for.

Gemini 3.7 Flash landed on August 13, three weeks after 3.6 Flash, which is a fast turn even by this industry's standards. Google is calling it a workhorse and pointing it squarely at two jobs: writing code and running agents. I have spent four days trying to make it fail at both, because that is the job, and because a workhorse is a thing you judge by Thursday, not by the launch chart.

The price is the headline, and the catch is a date

Start with the number, because Google did. The introductory price is $0.75 per million input tokens and $3.75 per million output — half of what 3.6 Flash cost at its own launch. For anyone running a model in a loop, which is what an agent is, that halving is not a rounding error. It is the difference between a background task you can leave running and one you watch nervously.

Here is the catch, and it is a good one. The half price is introductory, and "introductory" has an expiry: on January 1, 2027, the rate goes to $1.50 and $7.50, exactly double. So the model you are budgeting for in December is not the model you are paying for in January. If you are the kind of person who builds a workflow around a price and forgets it, this one has a fuse on it. Reaches-for-the-highlighter moment: whatever you prototype cheaply over the next four months, price it again at the standard rate before you promise anyone it is sustainable.

I mention this first because it colours everything else. A model that is very good and temporarily half price is a different product from a model that is very good. The discount is doing part of the persuading, and it is worth separating the part that is skill from the part that is a sale.

The coding jump is real. Google's numbers are Google's.

On paper the improvement over 3.6 Flash is not subtle. Google's own figures have it going from 34.4 to 43.6 percent on a production-code-quality benchmark, and from 49.0 to 65.3 percent on a long-horizon software-engineering test — the kind that measures whether a model can keep its head across a task with many steps rather than one clever function. Web-development scores are up too. Those are the reported numbers, and they are the numbers from the company selling the model, which is the correct amount of trust to extend to them: some.

So I stopped reading the chart and gave it work. Across the weekend I handed it four jobs from my actual backlog: the failing test, a small feature that needed a new endpoint and a migration, a refactor I had been dreading, and a documentation pass on code I wrote last year and no longer understand. On the first three, it was genuinely good — not demo-good, Thursday-good. The endpoint worked on the second try. The migration it wrote was cautious in the way a careful colleague is cautious, with a rollback path I did not ask for and was glad to have.

A workhorse is a thing you judge by Thursday, not by the launch chart. Four days in, this one earns the name — right up to the seam where the tasks touch.

The refactor is where it got interesting, in both directions. It did the mechanical part beautifully and fast. But it also made a judgment call — collapsing two similar functions into one "cleaner" version — that was defensible in isolation and wrong for my codebase, because those two functions are similar today and are going to diverge next month, and I know that and it cannot. That is not a bug. It is the exact shape of the thing you are supervising when you supervise a coding model: it optimises the code in front of it, not the code you are about to write.

"Agentic" is a promise about the boring parts

The other half of Google's pitch is agents, and this is the part every lab is selling right now, so it is the part worth being slowest to believe. 3.7 Flash ships with the usual agent kit — function calling, search as a tool, and computer use, the feature where the model drives a browser or a screen rather than just talking about one. Google's automation benchmark score nearly doubled, from 17 to 30 percent. Read that number the right way: doubling is real progress, and 30 percent is also the number of times out of a hundred it finished the automated task. Both halves of that sentence are true, and the second half is the one that will decide whether you can actually leave it alone.

I set it a small agent job that mirrors something I do most weeks: go through a folder of receipts, pull the amounts, and put them in a running sheet. One task. It did it, and it did it faster than I would have. Then I asked for the thing real work always asks for, which is the second task that depends on the first — reconcile those amounts against a list I already had, and flag anything that did not match. That is where it started to get human, in the bad sense. It read the receipts fine and it read my list fine. Stitching the two together — hold this while you check that — is where it guessed, and a guess wearing a spreadsheet cell is just a number you have to go back and verify by hand, which means you did the work anyway.

This is the same seam every agent frays at, and 3.7 Flash frays at it later than 3.6 did, which is the honest form of the compliment. The handoff between two tasks is still where the magic thins out. If your agent job is one clean action repeated — extract, classify, fetch — this model will save you real time. If it is two actions that have to know about each other, budget for a human at the join.

The context window is the quiet win

The least-hyped feature is the one that helped me most. 3.7 Flash keeps the million-token context window of its predecessor, and on the documentation job — the one where I no longer understood my own year-old code — I dumped the entire module, its tests, and a year of commit messages in at once and asked it to explain what past-me was thinking. It did, accurately, including catching a comment that contradicts the code it describes. Google reports a big jump on document comprehension, from 22 to 34 percent, and this was the one place the improvement felt less like a benchmark and more like a different tool.

That is the unglamorous case for this model. Not the agent that runs your life. The very large, very cheap reader that will hold a whole messy codebase or a whole contract in its head and answer questions about it without losing the thread. In practice that is the job I trust it with most, precisely because a wrong answer there is easy to check and cheap to ask again.

Who it's for, who should wait

So, the verdict, which is the part I have to live with. If you write code, or you are building the kind of agent that does one bounded thing at a time, 3.7 Flash is an easy yes for the next four months — it is clearly better than 3.6 at the coding work, and at half price the math barely needs doing. Prototype aggressively, and set a reminder for late December to re-price everything at the standard rate, because that is when the sale ends and the real cost shows up.

If your plan is a hands-off agent that chains several dependent steps and needs to be right without a babysitter, wait, or at least test it on your own two-task job before you believe the demo. The improvement is real and the seam is still there. I have added 3.7 Flash to my kept list, not my cancelled one — the first model in a while to earn that on coding alone. But I kept it for the reading and the writing, not for the promise that it will go do my week for me. That model keeps getting closer. This is not the Thursday it arrives.