AI agents cost between £0.01 and £5+ per task. The actual price depends on the model tier, the number of tokens consumed, the complexity of the task, and whether your workflow triggers retries or multi-step orchestration.
That range is wide — and deliberately so. A simple classification task using a small model might cost a fraction of a penny. A multi-step research workflow using a large reasoning model might burn through £5 or more in a single run. Most teams have no idea where their spend falls on that spectrum. Without granular tracking, AI agent costs become invisible line items buried inside cloud bills. Understanding the real cost of AI agents is the first step toward controlling it — and toward billing clients accurately for AI agent work.
What Do AI Agents Cost Today?
AI agent costs break down into three layers: token consumption, compute infrastructure, and orchestration overhead. Token consumption is the largest variable.
According to provider documentation, large frontier models charge roughly $2.50 per million input tokens and $10.00 per million output tokens. Smaller, faster models sit between $0.15–$0.30 per million input tokens and $0.60–$1.20 per million output tokens. The model you choose determines the baseline.
But the task determines how many tokens get consumed. Here is what typical workloads look like:
| Task Type | Token Range | Estimated Cost (Large Model) | Estimated Cost (Small Model) |
|---|---|---|---|
| Simple classification | 500–2,000 tokens | £0.01–£0.02 | < £0.01 |
| Email drafting | 2,000–8,000 tokens | £0.03–£0.08 | < £0.01 |
| Research summary | 15,000–50,000 tokens | £0.15–£0.50 | £0.01–£0.06 |
| Code generation | 50,000–200,000 tokens | £0.50–£2.00 | £0.03–£0.24 |
| Multi-step analysis | 100,000–500,000 tokens | £1.00–£5.00+ | £0.06–£0.60 |
These figures cover raw API costs only. They exclude orchestration, retries, and infrastructure — which can double or triple the effective price.
For a 10-person team running AI agents across daily workflows, industry surveys place typical monthly spend between $500 and $5,000. Teams without cost controls tend to land at the higher end.
How Are AI Agents Priced?
Four pricing approaches dominate the market. Most providers use one or a combination.
| Pricing Approach | How It Works | Advantage | Disadvantage |
|---|---|---|---|
| Per-token | Pay per input and output token processed | Pay only for what you use | Costs are unpredictable at scale |
| Per-task | Fixed fee per completed action | Predictable budgeting | Overpay on simple tasks |
| Subscription | Monthly seat or usage tier | Cost ceiling is known | Waste if usage is low |
| Hybrid | Base subscription + per-token overage | Balances predictability and flexibility | Complex to model internally |
Per-token pricing is the most common for direct API access. You pay exactly for the tokens your agent consumes. This gives maximum granularity but makes forecasting difficult. A workflow that runs cleanly one day might hit retry loops the next and cost five times as much.
Per-task pricing is growing among managed agent platforms. You pay a flat fee per completed task — a document review, a data extraction, a code fix. This shifts the cost risk to the provider. It also makes client billing straightforward, since each task maps to a line item.
Subscription models bundle a fixed amount of usage into a monthly fee. They work for teams with predictable workloads. They fail when usage spikes or drops significantly month to month.
Hybrid models combine a base subscription with per-token charges above a threshold. This is where most enterprise pricing is heading. It offers a cost floor for budgeting with flexibility for variable demand.
What Drives the Cost Differences Between Providers?
Not all providers charge the same rates for similar capability. Three factors explain the gaps.
Model tier. Large reasoning models with extended context windows cost 10–50x more per token than smaller, task-specific models. A coding agent that defaults to a frontier model for every step will cost dramatically more than one that routes simple sub-tasks to a smaller model.
Cloud platform markup. Accessing models through a major cloud platform typically adds 10–30% over direct API pricing, according to provider documentation. That markup covers managed infrastructure, compliance certifications, and support. For some organisations, those extras justify the premium. For others, direct API access is the better deal.
Open-source alternatives. Self-hosted open-source models eliminate per-token API charges entirely. The trade-off is infrastructure cost: GPU instances, maintenance, and the engineering time to keep models running. For high-volume workloads, self-hosting can cut costs by 50–80%. For low-volume use, the operational overhead rarely justifies it.
The right choice depends on volume, latency requirements, and internal engineering capacity. Most teams end up using a mix — large models for complex reasoning, small models for routine tasks, and self-hosted options for high-frequency workflows.
What Hidden Costs Do Most Teams Miss?
The API bill is the visible cost. The hidden costs often exceed it.
Orchestration overhead. Multi-step agent workflows require planning, routing, and state management. Each orchestration layer adds its own compute cost and latency. A five-step workflow does not cost five times a single step — the orchestration glue between steps adds 15–30% to total token spend.
Retry loops. When an agent fails a step, most frameworks retry automatically. A single retry doubles the token cost of that step. According to practitioner reports, poorly tuned agents can trigger 2–5x token multiplication through retry loops alone. A task budgeted at £0.50 can quietly cost £2.50.
Context window bloat. Agents that carry full conversation history into every step waste tokens re-processing old context. A 10-step workflow that passes the full history at each step processes far more tokens than one that summarises and truncates. Teams that do not manage context windows pay for the same information multiple times.
Monitoring and observability. Tracking what your agents do requires its own infrastructure. Logging, dashboards, alerting — these add operational cost. But the alternative — running agents blind — costs more in wasted spend and missed errors. Understanding what AI agent time tracking involves helps teams build the right monitoring layer without over-engineering it.
Human oversight. AI agents still require human review for high-stakes outputs. The cost of that review — the time a senior team member spends validating agent work — is rarely tracked or allocated. It should be.
How Can You Track and Control AI Agent Costs?
Cost control starts with visibility. You cannot manage what you cannot measure. Four strategies keep AI agent spend under control.
Set token budgets per task. Define maximum token limits for each task type. A research summary gets 50,000 tokens. A classification gets 2,000. If an agent hits the ceiling, it stops or falls back to a smaller model. This prevents runaway costs from a single misbehaving workflow.
Track cost per task, not just total spend. Monthly totals hide the detail. Tracking cost at the individual task level reveals which workflows are efficient and which are burning budget. This is the same principle behind AI cost management — granularity drives better decisions.
Route tasks to the right model. Not every task needs a frontier model. Build routing logic that sends simple tasks to small models and reserves large models for work that genuinely requires them. Teams that implement model routing typically reduce costs by 40–60% without measurable quality loss on routine tasks.
Integrate cost tracking with time tracking. When AI agent costs flow into the same platform as human time tracking, managers see total project cost in one view. This matters for client billing, internal budgeting, and ROI measurement. The connection between agent costs and billable hours for AI agent work is direct: every token spent on a client project is a cost that needs to be recovered or absorbed.
Key Takeaway
AI agents cost £0.01 to £5+ per task. Control spending with token budgets, model routing, and per-task cost tracking tied to your time tracking platform.
Get Full Visibility Into AI Agent Costs
Keito tracks every AI agent action — tokens, time, and cost — alongside your human team’s hours. See exactly what each task costs.
Frequently Asked Questions
How much does it cost to run an AI agent?
Running an AI agent costs between £0.01 and £5+ per task, depending on the model tier and task complexity. Simple tasks like classification cost fractions of a penny. Multi-step research or coding workflows using large models can exceed £5 per run. Monthly costs for a 10-person team typically range from $500 to $5,000.
What is the cheapest way to run AI agents?
The cheapest approach is to route simple tasks to small, low-cost models and reserve large models for complex work. Self-hosting open-source models eliminates per-token API fees for high-volume workflows. Setting token budgets per task prevents runaway costs from retries and context window bloat.
Do AI agent costs include infrastructure?
Per-token API pricing covers model inference only. It does not include orchestration, monitoring, retry costs, or human oversight. These hidden costs can double the effective price. Cloud platform access adds a further 10–30% markup over direct API pricing.
How do you budget for AI agent spending?
Start by tracking cost per task across your workflows for two to four weeks. Use that data to set token budgets per task type. Build model routing so expensive models only handle tasks that need them. Review spending weekly and adjust budgets as usage patterns stabilise.
Can you bill clients for AI agent costs?
Yes. Track agent activity at the task level — tokens consumed, time elapsed, and total cost — then present itemised breakdowns on invoices. Per-task or cost-plus pricing models work best. See the full guide on billing clients for AI agent work for detailed approaches.