Cheap Tokens, Expensive Agents: The Unit Economics of AI Agents

If you priced an AI project six months ago and haven't repriced it, the headline numbers have moved dramatically in your favor. A September 2026 roundup of model updates counted more than 20 releases in a two-week window and a new price floor of roughly $0.10 per million tokens. LLM Gateway's release timeline logged 23 new models from 15 providers in September alone. Frontier releases — GPT-6 Astra, Claude Opus 5.5, Gemini 3.8 Flash, Meta Muse — all landed inside the same few weeks.

So tokens are close to free. Why are agent pilots still blowing through budgets?

Because the unit you're buying isn't a token. It's a resolved outcome — a closed ticket, a reconciled invoice, a qualified lead, a completed intake. And an agent doesn't produce an outcome with one model call. It produces it with dozens: planning turns, tool calls, retries, re-read context windows, and reasoning-effort settings almost nobody configures deliberately. The price per token fell. The number of tokens per outcome went up faster.

The headline versus the invoice

Here's the arithmetic that surprises teams. Suppose you estimate a support resolution at 4,000 tokens of input and 800 tokens of output. At floor pricing, that's a fraction of a cent. Comfortable. You green-light the pilot.

Then the agent actually runs. It takes seven turns instead of one. Each turn re-reads the conversation history plus retrieved documents, so your 4,000-token context is re-sent seven times and grows as it goes. It makes four tool calls, two of which return errors and trigger retries. The reasoning-effort default shipped by the vendor is set to "high," so every turn burns hidden reasoning tokens you never see in your prompt. One in eight sessions escalates to a frontier model for a second pass.

Your fraction-of-a-cent resolution is now measured in dollars — sometimes an order of magnitude or two above the estimate. Nothing broke. The agent worked. The economics broke, quietly, in a place your pricing spreadsheet wasn't looking.

This is not a theoretical concern at the infrastructure layer either. Anthropic reportedly committed roughly $11.6 billion over seven years to Akamai's distributed cloud specifically for agent workloads, and xAI has signalled Colossus 2 may more than double its Nvidia chip count by the end of 2026. Read that as a market signal: agent inference, not training, is becoming the dominant cost center. The people closest to the compute are provisioning for it. Most enterprise buyers are not.

The 119x trap

The most useful number to come out of this month's pricing analysis is a 119x spread in per-task cost between models that post broadly comparable benchmark scores. Same capability tier. Two orders of magnitude difference in what it costs to finish the job.

That spread exists because benchmarks measure whether a model can do something. Your P&L measures how many tokens it burns doing it — how verbose its reasoning is, how many turns it needs, how efficiently it uses tools. A model that solves the task in three tight turns and a model that solves it in fifteen rambling ones can score identically on a leaderboard and differ wildly on your invoice.

The practical implication: model selection is a cost decision as much as a quality decision, and it can no longer be made once. With 23 models from 15 providers arriving in a single month, whichever model is cheapest-per-outcome for your workflow today probably isn't in 90 days. Routing shipped as a standalone product this month for exactly this reason — LLM Gateway launched its Smart Route router on September 25, 2026.

Four cost levers most teams never pull

• Tiered model routing. In most agent workflows, 70–85% of steps are mechanical: classify, extract, format, validate, look up. Those belong on a cheap, fast model. The genuinely hard reasoning steps — ambiguous judgment, multi-document synthesis, exception handling — get the frontier model. Teams that route this way routinely cut cost per outcome by a large multiple without touching accuracy, because the cheap model was never the bottleneck on quality.
• Reasoning-effort defaults. Vendors now ship "effort level" settings in their documentation, and the defaults are generally tuned for benchmark performance, not your budget. Leaving them untouched means paying maximum reasoning cost on tasks that need none of it. Set effort per step, not per application.
• Context discipline and caching. Re-sending the same system prompt, tool schemas and retrieved documents on every turn is where agent costs actually explode. Prompt caching, aggressive context pruning, and summarizing prior turns instead of replaying them attack the largest line item first.
• Loop ceilings and hard budget caps. Every agent needs a maximum turn count, a maximum retry count, and a per-task spend ceiling that fails over to a human or a fallback path when hit. Without it, a single malformed input can run a loop until someone notices the bill.

Why this is an architecture decision, not a procurement decision

Here's the part that matters for how you buy. Routing, caching, effort control and spend enforcement all have to live in your orchestration layer. A per-seat SaaS agent hides the economics entirely — you have no visibility into tokens per outcome and no lever to pull when the vendor's model changes underneath you. A thin wrapper around a single provider can't route at all, because there's nowhere to route to.

This is precisely the case for building agentic AI systems on an orchestration layer you own, with LLM integrations that treat the model as a swappable component. The same architecture that protects you from model churn protects your margins — one abstraction, two problems solved.

The production gap is a money problem

The adoption data tells a consistent story. Forrester's The State of Agentic AI in 2026 found roughly 75% of enterprise leaders say they're adopting agentic AI, while only a small minority run anything meaningfully in production beyond what the report calls "agentish" chatbots. Gartner's 2026 Hype Cycle puts deployment at 17% of organizations, with more than 60% expecting to deploy within two years. Meanwhile the enterprise agentic AI market is projected to grow from $3.67B in 2025 to $24.50B by 2030 — a 46.2% CAGR, per an August 2026 analysis synthesizing 20+ analyst reports.

Everyone is piloting. Almost nobody has scaled. The usual explanation is capability, but capability isn't the binding constraint anymore — a pilot that works on 50 tickets works on 50,000. What stops it is that nobody can answer the CFO's question: what does this cost at 50,000? If the answer is a shrug, the project stays a pilot.

The market is responding. Dataiku launched Agent Management on September 24, 2026, bringing permissions and audit to agent fleets, and OpenAI shipped a Data agent in ChatGPT Work on September 9. Governance and cost control are productizing — a reliable sign that the industry has hit the same wall.

Our build standard

Every agent we ship at Levels AI includes three things before it goes near production traffic:

• Per-task cost telemetry. Tokens, turns, tool calls and dollars attributed to each completed outcome — not aggregate monthly spend, which tells you nothing actionable.
• An explicit routing policy. Which steps run on which tier, with the criteria written down and re-testable when a new model lands next month.
• A hard spend ceiling per task and per workflow, with defined failover behavior.

The result is that cost per resolved outcome is a dashboard number your finance team can see before you scale, alongside quality metrics. That's what turns an interesting pilot into a funded program.

Talk to us about your agent economics

If you have an agent in pilot and can't say what a single resolved outcome costs, that's the gap worth closing first. We'll run a short agent cost audit on one workflow — map the turn and tool-call profile, model the cost per outcome at your projected volume, and identify which of the four levers moves your number most.

Get in touch with Levels AI to scope it. No slide deck required — bring your workflow and we'll bring the math.