How Much Does an AI Agent Cost in Production? Real Monthly Costs

How much does an AI agent cost in production?

Running an AI agent in production costs $3,200 to $13,000 a month once you count everything beyond the model call. Most teams budget for a quarter of that, because they only priced the tokens. The exact number depends on how autonomous the agent is, but the pattern holds: the LLM is rarely the biggest line item.

Why is "how much do tokens cost" the wrong question?

Ask a finance team what an AI agent costs and they'll quote a per-million-token rate. That number has been falling fast. Gartner projects that running inference on a 1-trillion-parameter model will cost GenAI providers over 90% less by 2030 than it did in 2025. If token price were the whole story, agent budgets should be shrinking.

They aren't. Gartner separately forecasts that AI inference costs per agentic workflow will increase more than fivefold through 2028, even as per-token prices keep dropping. The firm calls this the inference paradox: cheaper tokens don't help when a workflow burns through vastly more of them, wrapped in infrastructure that bills independently of the model call. Enterprise usage data backs this up. Ramp's spending figures show business token usage grew over 1,000% between January 2025 and April 2026, and Menlo Ventures found enterprise generative AI spend jumped from $1.7 billion in 2023 to $37 billion in 2025. Volume is exploding faster than unit price is falling, and the stack wrapped around every call is scaling right along with it. The result: enterprise AI inference makes up 85% of total AI budgets by some estimates, yet Cockroach Labs puts model inference at only about 20% of an agent's actual total cost of ownership. The other 80% lives in orchestration, retrieval, evaluation, and governance.

What does a single agent task cost in tokens?

Start with the baseline unit, because everything else multiplies off it. Gartner's analysis found agentic workflows require 5 to 30 times more tokens per task than a standard chatbot query, since an agent plans, calls tools, reads results, and often revises its own output before finishing. Anthropic's own engineering data puts multi-agent systems at roughly 15 times the token consumption of a normal single-turn chat. At Uber, that multiplier showed up as a real budget line: when Claude Code adoption jumped from 32% to 84% of its 5,000-engineer organization, per-engineer monthly API costs ran $500 to $2,000, and the company's entire annual AI budget was gone by mid-April.

Caching and routing only take the token bill so far. Cutting it further is its own discipline, and worth doing before you touch the infrastructure layer underneath it.

How much do AI sandboxes cost to run agents in?

Most production agents don't just call a model. They run inside a sandboxed compute environment so they can execute code, browse, or touch files without putting production systems at risk. That compute is billed separately from the model, and the billing model matters as much as the rate.

Headline sandbox pricing clusters around $0.0504 per vCPU-hour on providers like E2B and Daytona, with a 2 vCPU, 4 GB instance running roughly $0.05 to $0.34 an hour depending on whether the vendor bills wall-clock time, active CPU, or metered usage. At that rate, a 45-minute agent session costs four to twenty-five cents in raw compute. That sounds trivial until you multiply it by concurrency and uptime across an entire fleet.

AWS Bedrock AgentCore shows how fragmented this billing gets. Its runtime microVMs charge $0.0895 per vCPU-hour and $0.00945 per GB-hour, but that's just the base compute meter. Browser, Code Interpreter, Gateway, Identity, Memory, and Observability are each billed separately, so a single agent task can touch eight or more independently metered line items before you even count the model call. None of that shows up if you only budget for tokens.

Why can infrastructure out-cost inference at scale?

The gap between a sandbox's hourly rate and your actual bill comes down to architecture, and the spread is enormous. Northflank ran an identical workload across providers at 200 concurrent sandboxes and found a monthly cost of $7,200 on its own PaaS, versus $16,819 on E2B and Daytona, $24,491 on Modal, and over $35,000 on Fly.io Sprites, for the same shape of work. Running the same 200 sandboxes on bring-your-own-cloud infrastructure dropped the bill to $2,060 a month. That's roughly a seventeen-fold difference in what should be a commodity cost, decided entirely by who owns the infrastructure and how it's billed, not by how much the agent actually did. This is the same dynamic driving the markups built into frontier model APIs: renting convenience at every layer compounds fast.

What does it cost to govern agent identities and permissions?

Every agent needs credentials: API keys, service accounts, OAuth tokens, scoped permissions to the systems it touches. These are non-human identities, and they're multiplying faster than anyone is governing them. The Cloud Security Alliance found non-human identities now outnumber human users by an average of 45 to 1, reaching 144 to 1 in cloud-native environments, and only 15% of organizations feel confident they can prevent an NHI-based attack.

Microsoft now prices this governance gap directly rather than bundling it:

• Entra Workload ID Premium costs $3 per workload identity per month, covering lifecycle and risk detection for the service accounts and tokens an agent uses.
• Microsoft Agent 365, the dedicated agent governance and observability layer, lists at $15 per user per month on an annual commitment, and is not included in standard E3 or E5 licensing.

Multiply either of those across a fleet of agents each holding several credentials, and the permissions layer alone can rival the inference bill for a mid-sized deployment.

How much does observability and evaluation cost for AI agents?

You can't run agents in production without watching what they do, and that visibility is its own meter. LangSmith charges $0.50 per 1,000 base traces after its free tier, and Datadog's LLM observability tier covers 40,000 spans a month for free before Pro pricing starts at $160 a month. Those numbers look modest in isolation, but an agent fleet generating thousands of traces a day crosses free tiers fast. Evaluation compounds the cost further, since confirming an agent actually did the right thing usually costs more than the inference itself once you run hundreds of test cycles per release.

What does a production agent fleet cost?

CloudZero's analysis of production deployments puts total monthly spend, across tokens, vector database hosting, monitoring, prompt tuning, and security, in bands that scale with autonomy:

• Assisted agents, triggered by a human for a specific task, run roughly $150 to $600 a month.
• Semi-autonomous agents, operating with some independence inside guardrails, run $1,200 to $5,500 a month.
• Fully autonomous multi-agent systems run $7,000 to $15,000 or more a month.

Most production deployments land between $3,200 and $13,000 a month in total operational spend once they're past the pilot stage. As vendors shift from flat per-seat pricing toward metered, usage-based billing, that range is getting harder to forecast month to month, not easier.

Why do agent bills blow up unpredictably?

The numbers above assume stable, well-behaved operation. Agents don't always stay well-behaved, and that's where the real budget damage happens.

• Context rot and retries. As an agent's working context grows across a long session, it tends to lose coherence and re-run steps, multiplying token use without multiplying useful output.
• Retrieval faults cascading into cost. A healthcare company running a three-agent deployment saw monthly inference costs jump from $12,000 to $68,000 in six weeks after a single retrieval fault sent the agents into repeated, expensive retries.
• Runaway autonomous loops. Developer Peter Steinberger's OpenClaw agent burned $1.3 million in 30 days. On a smaller scale, a founder's nine-day "vibe-coding" session on an AI coding agent racked up $607.70 in overage charges in three days, right before the agent deleted a production database. The spend spike was the warning sign of bad behavior, not just a budgeting nuisance.

Gartner now predicts more than 40% of agentic AI projects will be canceled by the end of 2027, and the stated reasons are escalating costs and unclear business value, not model incapability. The models work. The bills around them are what kill the project.

How can you control the bill?

The fix isn't a cheaper model. It's controlling the multiplier.

1. Cache aggressively and route by task. Repeated context and simple sub-tasks should never hit a frontier model at full price. Cheaper models and cached reads can absorb a large share of volume without hurting output quality.
2. Cap loops and retries explicitly. Set hard ceilings on how many times an agent can retry a failed step before it escalates to a human, rather than letting it retry indefinitely.
3. Own the infrastructure where volume justifies it. The seventeen-fold spread between PaaS and bring-your-own-cloud sandbox pricing shows that at real scale, owning the runtime layer instead of renting it is often the single biggest lever available.
4. Price identity and observability into the budget from day one, not as a surprise line item discovered after the pilot.
5. Track cost-per-task, not cost-per-token, so a sudden multiplier shows up in the metric that actually predicts budget blowouts.

That last point is the one most teams skip, and it's the one that matters most. A token price is a vendor's number. A cost-per-task is yours, and it's the only number that tells you whether the agent is worth what it costs to run.

The token price was never the budget. It was the entry fee. The sandbox, the identity layer, the observability platform, and the retries you didn't plan for are the bill.