On September 29, 2026, Perplexity added Automations to its Computer agent. The feature lets an agent run on a schedule or wake up when something happens in Slack, Gmail, Outlook, Linear or GitHub. Buried in the launch details was one line that tells you where agent pricing is going. As AlphaSignal summarised the release, "Monitoring for triggers consumes no credits. Credits apply when Computer runs an assignment."
Eleven days earlier, on September 18, AWS shipped a new runtime for Amazon Bedrock AgentCore. The AWS post says customers now "pay only for the memory that your agent uses, loaded on demand and reclaimed when idle." Under the old runtime, allocated memory stayed on the bill until the session ended. Under the new one, an idle agent costs close to nothing.
On September 25, Microsoft introduced Copilot Autopilot. Jared Spataro, Microsoft's Chief Marketing Officer for AI at Work, described it as an agent you "give a name, a role and a goal," which then gets on with "watching channels, following up on threads, running recurring work and picking a project back up days later, without waiting for a prompt." Autopilot runs on usage-based billing.
Three companies, one month, one design choice. Waiting is now free and acting is metered. That one change moves the economics of AI agents to a place most finance teams are not watching. The model price used to be the big variable. Now the big variable is the trigger: the event that wakes the agent. Your company produces those events, and you have almost certainly never counted them.
The month the price of a task fell
The model releases in September all pushed the same story, and it is worth getting the numbers right before arguing with it.
On September 22, Anthropic released Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens. Anthropic's announcement says that "at default settings it will cost 40% less than Opus 5 on typical workloads." OpenAI released GPT-6 Sol and GPT-6 Luna the same day. MarkTechPost reported Sol at $2 input and $10 output per million tokens, half the previous tier's price. On AutomationBench 1.0.6, OpenAI said Sol scored "33.2% at $0.27 per task," while Claude Opus 5 at max effort scored 26.9% "at 11.1x that cost." Luna, at $0.10 and $0.50 per million tokens, reportedly cost "93% less per task than Opus 5" on a coding benchmark.
These launches did not lead with a per-token price. They led with a per-task price, and that is a real change in how labs sell. They now compete on the cost of getting a unit of work done, which is the number a buyer actually cares about.
Most executives will read those figures and conclude that agents are getting cheap. Per run, they are. The mistake is assuming the bill falls with the per-run price. The bill is the per-run price multiplied by the number of runs. September cut the first number and, through free watching, removed the one thing that held the second number down.
The economics of AI agents now lives in the event stream
Here is the arithmetic, kept deliberately simple.
The monthly cost of an always-on agent is roughly the number of events it watches, times the share of those events that make it fire, times the cost of each run. The vendors control the third term. You control the first two, whether you know it or not.
Under the old setup, a chatbot waited for a person to type. A person's patience limited the number of runs. Nobody asks a support bot four thousand questions a day. Watching agents remove that limit. Autopilot watches a Teams channel. A Perplexity automation watches a Gmail inbox. An AgentCore session sits in memory for nothing and wakes when its queue gets a message. The rate of runs is now set by how many messages, tickets, commits, emails and status changes your organization produces.
Here is an illustrative example, using a public benchmark price and invented volumes. An operations channel gets 400 messages a working day. Someone sets up an agent to "follow up on anything that looks like a blocker." The filter is loose, so the agent fires on a quarter of the messages. That is 100 runs a day. At Sol's AutomationBench figure of $0.27 per task, that one agent costs $27 a day, about $590 over a 22-day working month. That is modest. Now add an agent per team, per inbox and per repository, each built by a different person on a different afternoon. A company with 200 such agents is spending six figures a month, and not a single model price has gone up.
Zeus Kerravala, principal analyst at ZK Research, wrote in SiliconANGLE on September 28 that agentic workloads can use 10 to 100 times more tokens per task than a simple inference call. He cited a Futurum forecast of total inference spending rising from $120 billion in 2025 to $885 billion in 2030. In the same piece, Mazda Marvasti, CEO of Amberd.ai, described what happens when a useful tool spreads across an organization and is "now priced on a variable basis": the costs climb fast and the forecast breaks. Kerravala's point is that some organizations have dropped automation projects because they could not forecast what the projects would cost.
The forecasting problem has a specific cause. Finance teams forecast from things they already measure: seats, contracts, compute reservations. Nobody in finance measures event volume by channel. The cost driver lives in a place the budget has never looked.
Your noisiest channel is now a cost center
This leads to a strange conclusion. The chattiest parts of your company will become the most expensive to automate, and the reason has nothing to do with how hard their work is. They just make more noise.
A team that sends five status updates where one would do, copies twelve people on every thread and opens a ticket for every minor question will make its watching agents fire five times as often as a quiet team. The agent will read each update, judge most of them irrelevant and bill you for every judgment. Bad communication habits used to cost attention. Now they also cost money, at a price per event, every working day.
There is an upside. Cleaning up how a team communicates now has a measurable return. Before September, nobody could put a dollar figure on cutting a pointless weekly status email. Once that email wakes three agents across a hundred inboxes, you can.
Who writes the trigger owns the budget
Microsoft's pricing change matters most here, because Microsoft is where most companies will meet this problem first.
Keith Kirkpatrick of Futurum Group wrote on September 30 that Microsoft has split Copilot in two. "Everyday AI," meaning Copilot in Chat, Word, Excel, PowerPoint, Outlook and Teams, stays on a per-user subscription with fair-use limits. "Advanced AI," which covers long-running agents, unsupervised workflows like Cowork, Code and Autopilot, and frontier models, runs on Copilot Credits at $0.01 each. Kirkpatrick notes that consumption "varies with model, runtime, context, and tools." He sums up the structure as "seat for breadth, a meter for depth."
The seat pays for the person. The meter pays for whatever that person sets running. Autopilot's setup asks for a name, a role and a goal, which is about the same effort as writing a job ad, and anyone who can write a job ad can now create a standing cost.
For enterprise customers, Microsoft keeps usage-based billing switched off until an administrator creates a spending policy. For smaller customers, the defaults lean the other way: from October 19, 2026, new Microsoft 365 Copilot Business subscriptions bought direct will arrive with usage billing switched on, an Azure billing resource created automatically and a default spending policy applied. Microsoft says cost tracking for agents built in Copilot Studio will arrive in Agent 365 in October.
So Microsoft has provided a ceiling and a dashboard. Neither tells you whether a given trigger is worth firing. A spending cap stops the bleeding after the agent has already fired on ten thousand irrelevant messages. It does nothing about the design choice that let it fire in the first place.
Kirkpatrick also makes a point that should worry any CFO. Microsoft decides where the fair-use limits on the seat sit, and Microsoft can move them. When a vendor controls the line between flat-rate and metered work, it controls how much of your workload ends up on the meter. Each time the line moves, work that used to be included becomes billable.
Why vendors made watching free
Making the idle state free looks generous, but it is a sound business decision, and it helps to understand the incentive before you plan around it.
An agent that costs nothing while it waits is easy to say yes to. Nobody asks for budget approval to have something "keep an eye on the inbox." The purchase decision disappears at the moment it would normally happen, when the agent is created, and comes back weeks later as a usage line on an invoice. That is how cloud storage spread through companies a decade ago, and how mobile data plans worked before that. A free first step leads to a metered habit.
Free watching also gives the vendor something more valuable than a fee. The agent sits on your event stream. It reads every message in the channel to decide whether to act. Perplexity's Automations carry "memory of prior runs, so agents build on previous work instead of restarting." Autopilot has its own identity and memory inside your tenant. An agent that has watched your operations channel for six months knows a great deal about how your company works. Moving it to another vendor means starting that knowledge from zero. The free watching period is when the vendor builds the switching cost.
None of this makes the vendors dishonest. The prices are published and the per-task numbers really are falling. But the vendor's best outcome is a large number of always-on agents with loose triggers, and your best outcome is a small number of agents with tight ones. Those goals differ, and the default settings will reflect the vendor's.
The design work nobody has assigned
If the trigger drives the cost, then designing triggers is a financial job, and today nobody in most companies holds it. The engineer who wires up an AgentCore session thinks about latency. The team lead who names an Autopilot agent thinks about the goal. The finance team sees one total. Nobody is accountable for how often things fire.
There are four practical steps. None of them requires buying anything.
Keep a register of triggers, separate from your list of agents
Most governance work this year has focused on keeping track of agents: who owns each one, what it can access, which model it uses. SAP, ServiceNow and Microsoft have all shipped tools for this. An agent inventory answers the security questions. For cost, you need a trigger inventory: every event source that can wake an agent, how many events it produces per day, and what fraction of them cause a run. That is the table your CFO needs, and none of the September releases provides it out of the box.
Price the trigger before you approve the agent
The approval question used to be "which model should this agent use?" A better question is "how many times a day will this fire, and what is each run worth?" A trigger that fires 300 times a day to catch the two messages that matter is a bad design at any model price. Perplexity's own example points to the fix: a Gmail trigger that activates "only when a particular sender requests a decision." A narrow filter at the trigger costs nothing. Letting the model do the filtering costs a full run per event. Perplexity also caps scheduled automations at no more than one run per hour, which is a cost decision presented as a product limit.
Set budgets by fire rate as well as by spend
Spending caps are blunt. A cap of $5,000 a month tells you nothing until you hit it, and then it stops everything at once, including the runs that were worth paying for. A budget in fires per day per trigger behaves differently. It flags a rogue filter on day one, when the run count jumps from 40 to 4,000, long before the dollar cap trips. This is the same lesson cloud teams learned with autoscaling: watch the rate, because the total arrives too late.
Treat communication cleanup as a cost project
If loose communication habits wake agents, then tightening those habits saves money, and the savings can now be measured. Merging three status channels into one, ending reply-all threads and closing stale tickets all reduce event volume. Give that work an owner and report its savings the way you would report any other efficiency gain. Most companies have never had a financial reason to fix how they communicate. As of this month, they do.
The objection: prices keep falling
The strongest objection is simple. Per-task costs fell 40% at Anthropic and 50% at OpenAI in one week. If that continues, today's runaway trigger will be cheap next year, so why design for scarcity?
The answer is that every past fall in the price of computing increased total spending. Cheaper storage meant companies kept everything. Cheaper bandwidth meant they streamed everything. Cheaper agent runs mean more people set up more watchers with looser filters, because each one feels free. Kerravala's 10x to 100x token multiplier per agentic task is the start of that pattern. Futurum's forecast of a sevenfold rise in inference spending by 2030 assumes prices keep falling, and spending still goes up.
Falling prices also hide waste. When each run costs a fraction of a cent, nobody looks at the run log. That works until there are tens of thousands of watchers, at which point the waste is spread so thinly across so many triggers that no single one is worth investigating, and the total is still the biggest number on the AI invoice. Cheap runs make the problem harder to see without making it any smaller.
There is also an exposure problem that has nothing to do with price. A watcher that fires on every message in a channel also reads every message in that channel, and passes each one to a model. Loose triggers waste money, and they also send more of your internal discussion to more vendors than anyone signed off on. Kirkpatrick's warning about who controls the limits applies here too. Your event stream is the most detailed record of how your company works, and free watching means you are handing it over by default.
What this does to the shape of the company
The deeper change is to how work gets organized.
For fifty years, companies organized work around requests. Someone asks, someone else does. The cost of a request was a person's time, and managers rationed it through queues, priorities and saying no. Watching agents change that. They act on what happens, without anyone asking. Request volume is no longer the main cost driver. Event volume is, and most events are not requests. They are side effects: a status changed, a comment was added, a file was saved, someone replied "thanks."
A company whose agents react to events has to decide which events matter. That is the same judgment managers used to apply by hand when they chose which emails to read, now written down and charged per use. Companies that write that judgment down carefully will run thousands of agents at a cost they can predict. Companies that let every team lead write their own will find the CFO asking why the AI bill doubled in a quarter when headcount did not change and no new project started. The honest answer will be that people started talking more.
The architecture decision in front of you
Buying more agents does not solve this. Every vendor will happily sell you more watchers, and each will be cheap to idle and cheap per run. The problem sits between the products, in the event plumbing, the trigger filters, the fire-rate budgets and the cleanup of how your teams communicate. No vendor has a reason to fix that for you, since their revenue grows when the agents fire more often.
The work is architecture. Decide which event streams are allowed to wake an agent. Decide who can create a trigger and on what terms. Put the fire rate into the budget and treat cleaner communication as a cost saving. Do it now, while you have dozens of watchers rather than thousands, because once the knowledge in those agents' memories has built up, the cost of redesigning goes up with it.
Agor AI Advisory does this work with executive teams: we map the event streams, price the triggers and design agent systems whose costs a CFO can forecast. The vendors set the price per run. How many runs you pay for is your decision, and you should make it before the defaults make it for you. Schedule a strategic consultation with us today.
Sources
- Microsoft Official Blog, "Introducing the new Copilot with Home, Code and Autopilot," September 25, 2026
- Futurum Group, "Microsoft Splits the Copilot Pricing Model Into Everyday and Advanced AI," September 30, 2026
- AWS Machine Learning Blog, "The new AgentCore runtime," September 18, 2026
- AlphaSignal, "Perplexity's Computer Agent Now Automates Recurring Work Across Slack, Gmail, and GitHub," September 29, 2026
- Anthropic, "Claude Opus 5.5," September 22, 2026
- MarkTechPost, "OpenAI Releases GPT-6 Sol and Luna," September 22, 2026
- SiliconANGLE, "Agentic AI is breaking the token meter," September 28, 2026