Welcome to Unprompted: The AI Marketing Brief, where I cut through the noise in AI news and research to show marketers what’s happening — and why it matters for your work, your team and your career.

Can you help me solve this math problem?

"Jenny is a manager at a leading marketing agency. As part of an effort to conserve resources, she hands the new intern an important project. When the intern completes the assignment two weeks later, Jenny realizes that two-thirds of it is wrong. Jenny then spends her nights and weekends frantically fixing the project before the big presentation. How much money did Jenny save?"

The answer is ... nothing. Maybe less than nothing, once you count the stress and the wasted time.

It might sound ridiculous, or familiar, but this is how many marketing teams are approaching AI, especially agents. We've all heard that AI saves time and money. In an agent-first world, the reality is messier. Agents are far more sophisticated than traditional chatbots, and the work we hand them is bigger, with lots of interconnected steps.

Still, plenty of managers treat AI like an intern: cheap labor that can handle any project with a little direction. Then the results come back wrong, and the savings go out the window.

Yes, agents are getting cheaper per task. But many of the tools that handle real work are moving to pay-per-use pricing, which means our math problem might be coming for your 2027 budget.

So here's the question most marketers are skipping: What does it actually cost to hand work to an AI agent? In this edition of Unprompted, we'll look at how AI tools are changing and what those changes may actually cost you.

Microsoft Copilot's New AI Agents Come With a Meter

Let's start with an AI system that's built into the tools a lot of us use every day: Microsoft Copilot. With its latest update, Copilot lets users hand off complete deliverables like launch kits, RFP responses and customer briefings through a feature called Cowork. Microsoft also unveiled Autopilot, a persistent agent with its own identity and memory. It follows up on threads, runs recurring work, and contacts stakeholders without waiting for a prompt.

Sound too good to be true? In one way, it is. These longer-running agent tools use usage-based billing, separate from the flat per-user license that covers everyday Copilot use. That means the features most appealing to marketers, like delegating a launch kit or putting recurring work on autopilot, won't come with a fixed per-seat price. Teams that lean on them will need to track and budget for usage the way they would for ad spend.

AI Models Are Getting Cheaper: GPT-6 Sol and Claude Fable 5.1

That doesn't mean all AI is getting more expensive. OpenAI's new GPT-6 Sol and Luna models cost 50% less through the API than their GPT-5.6 versions. And the lower price doesn't seem to come at the expense of accuracy. On OpenAI's internal factuality test, Sol makes about half as many mistakes as its predecessor.

Not to be outdone, Anthropic released Claude Fable 5.1 and Claude Mythos 5.1. According to Anthropic, the two share the same underlying model with different levels of safeguards. Fable 5.1 is generally available, while Mythos 5.1 is limited to trusted-access programs. The price cut here comes from cache reads, which now cost 75% less. Anthropic says that cuts typical workload costs by about 25% compared to Fable 5, and up to about 45% for highly agentic work.

This might sound like good news until you realize that cheaper per task doesn't mean cheaper per finished task.

On AutomationBench, a Zapier benchmark of end-to-end workflows across sales, marketing, operations and other business functions, GPT-6 Sol scored 33.2% at $0.27 per task, according to OpenAI. By OpenAI's figures, Claude Opus 5 scored 26.9% at 11.1 times the cost.

So yes, the price drop makes AI more affordable for multi-step marketing ops work. But the best score on that benchmark still means about two out of three tasks didn't get done. Sound familiar? That's Jenny's intern. You still need a person checking and finishing the work, even if you have an "automated workflow."

Project Opal Shows What Handing Work to an AI Agent Looks Like

Finally, there's Project Opal. This Microsoft Copilot capability completes tasks in the background, using a secure Windows 365 Cloud PC to work across an organization's browser-based apps and data. According to Microsoft, it can handle jobs like managing security group requests, compiling audit evidence, triaging IT tickets and building financial close decks from Excel data.

You're not handing it the keys and walking away, though. Users can take control, pause, resume or guide the agent at any point while it works. And it has limits. Opal can run up to three jobs at once, and starting a fourth automatically closes out the oldest one. For now, it's only available to companies in Microsoft's Frontier early access program.

So what does all this mean for you? Add it up, and you get three costs: the bill, the price per task and the hours someone spends fixing what comes back.

The teams that win won't pick the cheapest model. They'll know which tasks come back clean and only meter those.

Things to Do This Week:

• Email your Microsoft 365 admin three questions. Is your org in the Frontier program, is usage-based billing turned on and what spending limits would apply to marketing?
• Run the three-times test. Pick one repetitive multi-app task, like pulling campaign results into a weekly report. Run it through an agent three times and count how many come back right without fixes.
• Check your API usage against the new pricing. If your team uses the Claude API or Claude Code for high-volume work, estimate what moving to Fable 5.1 saves.
• Write down your "clean list." Start a short list of tasks that agents finish reliably and tasks that still need a person. That list is your budget plan once metered billing arrives.