Building an AI agent pricing strategy that works

Why token rates alone will not tell you what an agent costs

An AI agent pricing strategy has to survive weeks like the one we just had. Three model providers changed their prices in the space of about ten days at the end of August and into September 2026. Google shipped Gemini 3.8 Flash, its third budget model in six weeks. Anthropic launched Claude Fable 5.1, up to 45 percent cheaper for agentic work than its predecessor. Meta released Muse Spark 1.3, undercutting the field at $0.55 per task. If you are trying to lock in a number for what your agents will cost next quarter, you are chasing a moving target. Build the strategy on how you measure cost, not on which vendor's rate card looks lowest this week.

We build agent systems for a living. The question clients ask us most is how much a given workflow will cost them three months from now. The rate card everyone quotes at us answers a different question, and it is the least useful part of the answer.

Per-token pricing hides what a "harder working" model actually spends

Google says Gemini 3.8 Flash "works harder" than its predecessor by taking more reasoning steps and calling tools iteratively on complex tasks. It launched at the same headline price as Gemini 3.7 Flash: $0.75 per million tokens. Same rate, different bill. Because the model burns roughly 30 percent more output tokens per task to do that extra reasoning, the effective cost per completed task goes up even though the price per token did not move at all.

Run the arithmetic on a single task. A workflow that used 10,000 output tokens on Gemini 3.7 Flash cost about $0.0075 at $0.75 per million tokens. The same workflow on 3.8 Flash, burning 30 percent more tokens to do its extra reasoning, costs closer to $0.0098. That is still far below Muse Spark's flat $0.55 per task, so the budget-model tier is not in danger here. The point is narrower: a 30 percent swing showed up with zero change to the advertised rate, and a team budgeting off last month's token count would have missed it entirely.

This is the trap in most pricing conversations we have with clients. A vendor announces a rate cut or holds a price steady, and everyone reads that as the story. The real number is the token count per task, which the rate card never shows you. A model that reasons more, calls more tools, or retries more often can cost you more per outcome while looking cheaper on paper.

What the current market actually tells you about agent economics

Line up the last few weeks of releases and a pattern shows up. Anthropic is competing on efficiency for agentic workloads specifically, rather than on raw price: Claude Fable 5.1 runs about 25 percent cheaper than Fable 5 typically, and up to 45 percent cheaper on agent tasks. Meta is competing on sticker price: Muse Spark 1.3 at $0.55 per task is the cheapest option in its class, though independent benchmarks from Artificial Analysis still put it behind Claude Fable 5.1 on capability. Google is competing on both fronts at once and occasionally working against itself, since a model that reasons harder can cost more even at a flat rate.

These are different competitions. A provider chasing the lowest per-task number and a provider chasing the best result per dollar on multi-step agent work are optimizing for different customers. If your workload is a handful of straightforward lookups, the cheapest per-task rate probably wins. If your workload is a multi-step process with tool calls, retries, and judgment calls along the way, the model that does more per call at a slightly higher token rate can end up cheaper overall, and more reliable.

Price the task, not the token

The fix is simple to state and takes real discipline to hold to: measure and price cost per completed task, not cost per token or per API call. Take the actual workflow, run it end to end, and add up everything it actually costs: the API calls, the retries, and the failed attempts that had to be redone. That number is the one that survives a vendor's next announcement.

Picture an ops lead who spends 6 hours a week reconciling supplier invoices by hand. The question worth answering before you touch a model is what one completed reconciliation costs end to end, and whether that number beats 6 hours of that person's time doing it by hand. Once you have that baseline, swapping the underlying model when a cheaper or faster one launches becomes a five-minute comparison instead of a re-negotiation.

Three things keep that comparison honest. First, build a fixed test set of real tasks pulled from your own workflow, not a vendor's benchmark, since benchmarks reward whatever the vendor optimized for and that may not match your workload. Second, track a cost-per-completed-task number over a real working week rather than a single demo run, because reasoning-heavy models behave differently under load than they do in a one-off test call. Third, put a recheck on the calendar every quarter at minimum rather than treating the pricing decision as final. Given that three major providers moved on price in the same ten-day stretch, a decision made in isolation goes stale fast, and switching cost between providers is itself part of the price you are paying.

Where this fits if you resell agent-powered work

If you are building an agent-powered feature that you charge your own customers for, the same logic applies one level up. Price your feature on the value of the completed task to your customer, then treat the underlying model cost as a margin line you watch, separate from the number you charge. A 30 percent swing in your model's token burn should not force a customer-facing price change every time a provider ships an update. Build enough margin at the start that a model swap stays an internal decision rather than a renegotiation with your customer.

This is also where scope matters: if your agent workload is genuinely simple, one well-scoped call per task, low retry rate, no multi-step tool chaining, a full evaluation framework may be more than you need. A single cheap model run against your real tasks for a week will tell you what you need to know. Save the fuller framework for workloads where the reasoning steps and retries actually add up to real money.

If you want help setting up that kind of measurement before you commit to a provider or a price, that is the kind of scoping work our AI services team does directly. We have watched this market move three times in ten days and built the habit of testing against real tasks rather than rate cards for exactly that reason.