Tool of the Week: Gemini 3.7 Flash
Google shipped Gemini 3.7 Flash on August 13. Two things about it matter, and neither is a benchmark.
It costs half of what the last one did. Introductory API pricing is $0.75 per million input tokens and $3.75 per million output, exactly half the launch price of 3.6 Flash. Those intro rates hold until January 1, 2027, when standard pricing ($1.50 / $7.50) takes over. So there is a real window here, and a real date on which your bill doubles if you build something and forget about it.
It got meaningfully better at the thing small businesses actually want: doing tasks, not writing essays. The agent and coding scores moved a lot in one release. DeepSWE went from 49.0% to 65.3%. AutomationBench, which measures exactly the multi-step "go do this across tools" work, roughly doubled from 17.0% to 30.4%. WebDev Arena Elo went 1538 to 1588. Read the AutomationBench number twice: doubling is a lot, and 30.4% is still not a number you'd bet an unsupervised process on. Both facts are true at once.
What it keeps: a 1 million token input context, 64k output cap, multimodal input across text, image, video, audio and PDF, function calling, search as a tool, and computer use. It is live now in the Gemini API, AI Studio, the Gemini app, Gemini Spark, Google Antigravity, Android Studio and Gemini Enterprise, across more than 160 countries. You do not have to wait for access.
One detail worth noticing: Google says it did not train this from scratch. It replaced the previous version using algorithmic improvements and user feedback. That is becoming the normal shape of model releases. Less "new brain," more "same brain, tuned on what people actually did with it."
What I'd actually do with this:
If you have anything running on 3.6 Flash, price out the swap this week. Same context window, same modalities, better task scores, half the cost. That is an unusually clean upgrade. Just put January 1, 2027 on your calendar so the price change is not a surprise.
Use it for volume, not for judgment. Flash-class models are the workhorse tier: classify, extract, summarize, route, draft. Send the small number of high-stakes calls to a bigger model. Most people overpay by running everything through the most expensive option they know.
Retest the automation you gave up on. If something failed six months ago because the model kept losing the thread across steps, that is precisely the axis that moved. Rerun your old failing case before you rebuild anything.
Do not read 30.4% on AutomationBench as "ready to run unattended." It is better, not solved. Keep a human on anything irreversible.
Who this is for: anyone already paying for AI on a per-token basis, or anyone who priced out an idea earlier this year and shelved it. The floor moved. The work is to go back and check whether your specific idea was under it.
Quick Hits
OpenAI previewed a tier that runs 14 times faster, and speed turns out to be a feature. Cerebras announced on August 13 that it powers GPT-5.6 Sol on a new OpenAI service tier called Ultrafast, generating up to 750 output tokens per second, up to 14 times faster than Standard processing. It is a limited API preview for selected customers, expanding as capacity grows. Early access went to Jane Street, Podium, Basis and Rogo. OpenAI did not publish a price, a region list, or an uptime commitment. Why it matters: you probably cannot get this yet, so treat it as a signal rather than a tool. The signal is that a whole category of work was blocked by response time, not by intelligence. A support agent that has to search three systems while a customer is still on the phone does not work at 30 seconds and does work at one second. If you have ever abandoned a live chat or phone answering idea because the pause felt wrong, that constraint is on its way out.
ChatGPT Ads is now in the UK, Japan, Brazil, Mexico and South Korea, with 31 European countries announced next. As of August 11, ads are running in those five markets, and OpenAI says a European expansion follows. Ads only appear for people on the Free and Go plans; Plus, Pro and Enterprise stay ad-free. Buying is currently through OpenAI's Ads Solutions team, agency partners and technology partners, with self-service access through an Ads Manager expected later. Why it matters: a new ad channel is opening, and early ad channels are historically cheap before everyone shows up. That is worth knowing even if you never buy one. It also tells you something about where people are now looking things up. If customers are asking an assistant instead of a search box, being findable there is a real question for the next year.
The open-weight side kept pace. Z.AI released GLM-5.2 Turbo on August 17, the newest member of a family whose base model, GLM-5.2, is a 753 billion parameter open-weight release under an MIT license with a 1 million token context window, out since June. Why it matters: open weights under a permissive license means you can run it on your own infrastructure, with your own data, with no per-token bill and no vendor able to retire the model out from under you. For most small businesses that is more effort than it is worth today. But it is the reason the paid models keep getting cheaper, and it is a genuine option if you have data that cannot leave your building.
xAI shipped always-on agents that get their own computer, and I want to see the bill. Grok Bot launched in beta on August 11. Each agent gets a persistent cloud computer with a browser, filesystem and terminal, signs into your existing tools with your credentials, and works multi-step jobs end to end while your machine is off, surfacing only when it needs approval. You cannot buy it on its own. Access is bundled into Cursor Teams Premium at $120 per seat per month, Cursor Ultra at $200 a month for individuals, or SuperGrok Heavy at $300 a month. Each plan includes some weekly usage, after which on-demand work is billed at model and token cost, and I could not find a published quota figure anywhere. Why it matters: two things stand out and neither is the headline. First, an included allowance with no stated cap, on a product designed to run while you are asleep, is a budget you cannot forecast. Ask what a bad week costs before you turn one loose. Second, it is ten days old and in beta, with no independent reliability record yet, which is exactly when the demos look best. I run a multi-agent setup myself, so I am genuinely interested in this one, and I am going to watch it for a few months before I tell you it works. If it holds up, you will hear about it here with real numbers attached.
Prompt of the Week: The Right Model for the Job
Most people pick one AI model and run everything through it. That is usually either too expensive or not good enough, and often both on different tasks. This prompt sorts your work into tiers before you spend anything.
I want to figure out which AI model tier each of my recurring tasks
actually needs, so I stop overpaying on simple work and stop
under-resourcing the hard stuff.
Here are the tasks I want to run with AI:
[list them: e.g. sorting inbound email, drafting quote follow-ups,
summarizing call notes, reviewing a contract, answering the phone]
For each task, ask me anything you need to know, then tell me:
1. Does this need a FAST/CHEAP model or a SLOW/SMART one? Explain in
one sentence which property the task actually depends on.
2. Is this task LATENCY-SENSITIVE, meaning a human is waiting on the
answer in real time? Say yes or no and why.
3. How bad is a wrong answer here, and is it reversible?
4. Based on 1 to 3, what should I actually do: run it on a cheap
model, run it on an expensive model, run cheap with an expensive
model reviewing, or not automate it at all yet?
Be blunt. If a task should not be automated right now, say so and
say what would need to change first.
Run it once on your five most repetitive tasks. The usual result is that three of them belong on the cheap tier, one belongs on the expensive tier, and one should not be automated yet. Knowing which is which is most of the savings.
One last thing
Like what you're reading? Forward it to one person who'd get something out of it.
And if you want a second set of eyes on where AI could actually save you time, book a free 15-minute audit.