Cheap tokens did not make agents cheap, and that gap is now the whole market. Through the first half of 2026, model prices kept falling while agent programmes kept overshooting their budgets — because the bill for an agent was never the price of a token. It is the price of a completed unit of work, including the retries, the tool calls, the retrieval, the orchestration overhead and the human review that remains. Buyers have caught on, and the questions they ask in Q3 2026 are harder and better than the ones they asked a year ago.
This article covers the three ways agents are priced today, why the sticker price misleads in both directions, and the evidence buyers now demand before signing. It is written for the person who has to defend the number to a finance team.
Refer & earn: 30% of a client’s first month, then 10% of every month after that, for up to 24 months per client. See the partner programme →
The cost curve is real, and it is public
Start with what is genuinely verifiable, because it explains half the optimism. Model prices are published, and the spread between a small model and a frontier model is roughly two orders of magnitude. As of the current published rate cards, Anthropic lists its mid-tier Claude Sonnet 5 at $2 per million input tokens and $10 per million output tokens, with cache hits at $0.20 per million, and its higher-tier Claude Opus 5 at $5 and $25 respectively. OpenAI’s published table spans from a low-cost model such as gpt-5.6-luna at $0.20 input and $1.20 output per million tokens, through gpt-5.6-terra at $2.00 and $12.00, up to a flagship in the same table, gpt-6-astra, at $10.00 and $50.00.
The lesson is not “AI is cheap”. It is that the choice of model is a first-order cost decision you are now expected to make deliberately. Ten times the input price buys a different class of reasoning; whether the task needs it is a question about the job, not about prestige. Teams that route routine work to a small model and reserve a frontier model for genuinely hard cases routinely spend a fraction of what teams that send everything to the flagship spend — for indistinguishable outcomes on the routine work.
Why the token price is not the bill
An agent is not a single model call. It is a loop, and each turn of the loop has a cost that does not appear on a rate card.
- Tool calls. Every external lookup, database query or API call is part of the work — and often part of the bill if the tool is metered.
- Retries and iteration. Anthropic’s engineering guidance is explicit that agentic systems trade latency and cost for task performance. That trade is not a bug to be fixed; it is the design. You pay for the thinking, not just the answer.
- Retrieval and context. An agent grounded in your documents spends input tokens on them every time. Grounding is what makes an agent safe and auditable, and it is not free.
- Orchestration overhead. When several agents collaborate, the hand-offs, the shared state and the merge steps all cost tokens and time. More agents is not always more output.
- Human review. The most under-modelled cost of all. Someone reads the drafts, resolves the escalations and supervises the agent. If supervision costs more than the task saved, you have learned something valuable early — see our piece on measuring an AI employee for how to count it.
Put those together and the honest unit of comparison is cost per resolved task, all-in: model, tools, orchestration and the human time still required. It is the number that lets you compare an agent to a person, and the only one a finance team should be asked to approve.
The three pricing models buyers are quoted
Vendors have converged on three ways of charging. Each has an honest use and a failure mode.
- Seat or licence. A flat fee per user or per agent per month. Simple, predictable, easy to budget — and disconnected from value. It rewards the vendor for provisioning, not for working, so it suits tools rather than workers.
- Usage pass-through. You pay for what the agent consumes, usually with a platform fee on top. Aligned with cost but not with outcome: an inefficient agent that burns tokens on a task it ultimately fails costs you the same as one that succeeds.
- Outcome or unit-of-work. You pay per resolved task, per completed document, per qualified lead — with a defined standard for “resolved”. The most honest alignment of incentives, and the hardest to define, because it forces both sides to agree on what counts as done.
The trend in 2026 is movement from the first toward the third, and it is driven by buyers. Once a procurement team understands that the token price is not the bill, a flat seat fee looks arbitrary. The counter-move from vendors is to keep the seat model but attach a performance guarantee — which is only credible if someone is measuring the performance.
Proof: what buyers now demand
The single biggest change since 2025 is that “show me a demo” has been replaced by “show me the evidence”. Demos run the two or three cases the builder knows work; production runs the distribution. Buyers have learned the difference, and the requests are now specific.
- A frozen test set of real tasks with known answers. Fifty real cases drawn from history, not imagination, with the answers authored by someone who knows the job. Re-run after every change.
- Success rate, escalation rate and cost per resolved task from launch. Not a projection. The numbers you will see in month one, reported the same way in month six.
- Failures kept visible. A vendor who reports only successes is telling you about their marketing, not their system. The failure list is the most informative document in the deal.
- References from a comparable deployment. Same shape of job, same class of organisation.
- Defined pilot exit criteria. Written before the pilot starts: what number, at what threshold, by when. Without this, a pilot can only be evaluated by sentiment.
- An audit trail a human can read and a tested off-switch. If the vendor cannot show you how to stop the agent instantly, that is the risk, whatever the pricing page says.
The shift is healthy, and it is what makes outcome pricing possible at all. You cannot pay per resolved task unless you can define and measure “resolved”. Enforcement questions matter here too — an agent that reads untrusted content and holds write access is a governance exposure as much as a cost one, which is why our guide to AI agent security sits alongside this one.
A buyer’s checklist for the Q3 2026 market
- Ask for the unit of work, not the unit of tokens. “What will this cost per resolved task, all-in?” is the only question with a defensible answer.
- Ask which model runs which step. A vendor who sends everything to the flagship is either not optimising or not telling you why.
- Ask for the supervision cost. How many human hours per week, at steady state? If the answer is “none”, be sceptical.
- Ask for the test set and the last three runs. The results should move when the model changes.
- Ask what happens when it is wrong. Escalation path, notification, rollback.
- Ask what “off” looks like. Then ask them to show you.
- Write the exit criteria before the pilot. A pilot without a threshold is a subscription with extra steps.
How Faida Workforce prices and proves its work. Faida Workforce builds AI employees as bounded specialists and reports success rate, escalation rate and cost per resolved task from launch — the numbers our clients take to their own finance teams. We prefer the unit-of-work conversation to the seat-licence one, and we would rather tell you an agent is not yet ready than sell you a demonstration. If you want the measurement plan before the build, start at faidaworkforce.com.
What to expect next
The direction of travel is stable, even if the prices are not. Model costs will keep drifting down and the price of the hardest reasoning will keep a premium; that is normal for infrastructure. What will not fall by itself is the cost of getting the work done — the tool calls, the retries, the supervision and the measurement. So the buyers who will look smart in 2027 are the ones doing two things now: insisting on cost per resolved task as the unit of comparison, and insisting on evidence rather than enthusiasm. The market has already rewarded them with better pricing models. It will reward them again with systems that survive their second budget review, which is a far better test than any demo.
Sources
- OpenAI — API pricing (per-million-token rates for current models; figures retrieved 15 September 2026): platform.openai.com/docs/pricing
- Anthropic — Claude platform pricing (per-million-token rates and cache pricing; figures retrieved 15 September 2026): platform.claude.com/docs/en/about-claude/pricing
- Anthropic Engineering — “Building Effective AI Agents” (agentic systems trade latency and cost for task performance; simplest-solution guidance): anthropic.com/engineering/building-effective-agents
- OWASP Gen AI Security Project — “Agentic AI: Threats and Mitigations” (excessive agency and related agentic risks): genai.owasp.org/resource/agentic-ai-threats-and-mitigations
- National Institute of Standards and Technology — AI Risk Management Framework: nist.gov/itl/ai-risk-management-framework
Illustration: Faida Workforce. This article is editorial content published under our editorial policy, including our disclosure on the use of AI in production.
30% of their first month, then 10% every month after that
Refer a business to Faida Workforce and you earn 30% of the client’s first paid month, then 10% of every month for up to 24 months per client — rising to 15% from the accrual after you pass 10 active clients. Paid by PayPal, batched on the 1st.
10 Pro clients = $2,598.70 over 24 months — $3,748.13 at the 15% tier.
Figures are potential earnings worked from the published rates, assuming every referred client pays for the whole cap — not a promise, not an average. The $1 trial is never commissionable.
