Agentic AI Token Costs: Why Your Enterprise AI Budget Is About to Explode

Uber's CTO reportedly told staff in early 2026 that the company's AI budget had been consumed faster than projected, with token costs from agentic workflows cited as the primary driver. That detail, surfaced by reporting from The Information and Axios, is the clearest public signal yet that enterprise AI finance is operating on the wrong model.

Most 2024–2025 AI budgets were built around seat licenses and conversational usage: a known number of users, a predictable interaction pattern, a manageable cost per seat. Agentic workflows break every one of those assumptions. The shift from chat to agent isn't a UI upgrade or a feature addition. It's a cost architecture change that invalidates the financial models finance teams spent the last two years building.

What 'Agentic' Actually Means for Token Consumption

An agentic workflow isn't a smarter chatbot. It's a reasoning loop: the model receives a task, decides what tools to call, retrieves context from memory or external sources, evaluates the result, and either completes the task or spawns a sub-agent to handle a component. Each of those steps is a separate LLM call. Each call re-sends the full context window accumulated up to that point.

That last sentence is where the math turns ugly. In a standard chat interaction, tokens accumulate linearly across a conversation. In an agentic loop, the context window is re-injected at every step. If your agent has a 32,000-token context and runs 10 reasoning steps, you're not sending 32,000 tokens total — you're sending something closer to 320,000 tokens, plus the output tokens generated at each step, plus tool call overhead.

The practical multiplier between a single conversational exchange and a multi-step agentic task can range from roughly 100x to 10,000x depending on context window size, loop depth, and whether sub-agents are spawned. A 500-token customer support reply and a 10-step invoice validation agent handling the same underlying question are not comparable billing events.

OpenAI's Nick Turley observed that unlimited AI plans are structurally similar to unlimited electricity plans — a useful analogy that exposes exactly why flat-rate pricing collapses under agentic load. Electricity providers don't offer unlimited plans because consumption is physically bounded. Token consumption in agentic systems is not.

Anthropic's Billing Pivot Is the Canary

In April 2026, Anthropic shifted enterprise customers from flat-rate seat pricing to a hybrid model combining per-seat fees with consumption billing at API rates. This wasn't a pricing strategy refinement. It was a structural acknowledgment that Anthropic's own usage data had confirmed what the math predicts: agentic customers consume tokens at a rate the flat-rate model cannot sustain.

Anthropic's internal data would have shown the distribution clearly. A small percentage of enterprise customers, specifically those running agentic workflows, were generating a disproportionate share of total token consumption. Flat-rate pricing socializes that cost across all customers. At some point, the math stops working.

The precedent matters beyond Anthropic. OpenAI and Google are watching the same usage data. Enterprise contracts negotiated in 2024 on chat-model assumptions will face renegotiation pressure as agentic adoption scales. Any enterprise that locked in a flat-rate deal expecting it to cover agentic usage should read Anthropic's move as a warning about what their next renewal conversation will look like.

Where the Costs Actually Accumulate: The Orchestration Layer

The billing events that compound agentic AI token costs don't all happen at the model API. They happen in the orchestration layer, and that's where most enterprises have the least visibility.

Tools like n8n, Make, and Zapier are widely used to wire together agentic workflows. Each workflow node that calls an LLM is a billing event. Visual workflow builders are designed to abstract complexity, which is useful for non-technical buyers but creates a significant blind spot: the person designing the workflow often has no direct line of sight to the token cost each node generates.

Mistral's Workflows layer, launched May 1, 2026, bundles orchestration directly with model access. That integration has real advantages for deployment speed, but it makes cost attribution harder to isolate. When the orchestration and the model billing are unified in a single product, it becomes more difficult to audit which workflow steps are driving consumption.

Voiceflow, a purpose-built platform for designing and deploying AI agents, illustrates a different cost dynamic. In a multi-turn agentic conversation, each turn carries forward the accumulated context of prior turns. A 20-turn customer service interaction isn't 20 independent LLM calls with 20 independent context windows. The context grows with each turn, and the token cost per call increases accordingly. Voiceflow scored 6.8/10 by the TopReviewed AI panel, and it's worth noting that its pricing model, like most platforms in this category, doesn't automatically surface per-conversation token costs to the workflow designer.

The Worked Cost Model: Chatbot vs. Agentic Workflow

Scenario A: Conversational Chatbot

A customer support chatbot handles 1,000 daily conversations. Each conversation averages 5 turns. Each turn averages 300 tokens of input and 200 tokens of output. That's 500 tokens per turn, 2,500 tokens per conversation, and 2.5 million tokens per day across 1,000 conversations. Monthly volume: approximately 75 million tokens.

Scenario B: Multi-Step Agentic Workflow

An invoice processing agent handles the same 1,000 daily tasks. Each task triggers between 8 and 15 LLM calls. At each call, the agent re-injects a 16,000–32,000 token context window covering prior steps, retrieved documents, and tool outputs. Using a conservative estimate of 10 calls per task at 20,000 tokens per call, that's 200,000 tokens per task. Daily volume: 200 million tokens. Monthly volume: approximately 6 billion tokens.

The table below presents these scenarios side by side. Costs are estimates based on publicly published API rates as of mid-2026 and will vary by model tier, prompt caching behavior, and whether batched inference discounts apply. Do not treat these as precise projections — treat them as order-of-magnitude indicators.

Scenario	Daily LLM Calls	Avg Tokens/Call	Monthly Token Volume	Est. Monthly Cost (Anthropic Claude 3.5 Sonnet)	Est. Monthly Cost (OpenAI GPT-4o)	Est. Monthly Cost (Gemini 1.5 Pro)
Conversational Chatbot (1,000 convos/day, 5 turns, ~500 tokens/turn)	5,000	~500	~75M tokens	Low hundreds of dollars (per published Anthropic API rates)	Low hundreds of dollars (per published OpenAI API rates)	Low tens of dollars (per published Google AI pricing)
Agentic Invoice Processing (1,000 tasks/day, 10 LLM calls, 20k context/call)	10,000	~20,000	~6B tokens	Tens of thousands of dollars	Tens of thousands of dollars	Thousands to low tens of thousands of dollars

The same task count, 1,000 daily items, produces a cost difference of roughly two orders of magnitude. That gap is not a rounding error in your AI budget. It's a structural mismatch between what was planned and what gets billed.

Why Google Vertex AI and Managed Platforms Don't Automatically Solve This

The instinct to move agentic workloads onto a managed platform like Google Vertex AI (scored 8.2/10 by the TopReviewed AI panel) is understandable. Managed platforms offer infrastructure abstraction, enterprise support tiers, and consolidated billing. None of that eliminates per-token charges.

Vertex AI's agentic features, specifically Agent Builder and grounding capabilities, introduce additional API calls per agent turn on top of the base model cost. Each grounding call that retrieves context from a data store is a billable event. Each tool invocation logged through the agent framework adds to the total. The platform markup on top of base model costs varies by configuration, but it doesn't trend toward zero.

The more insidious problem is the enterprise contract structure. Volume discounts negotiated in 2024 were modeled on chat-scale consumption projections. An enterprise that committed to a volume tier based on conversational usage will hit that tier's ceiling quickly once agentic workflows go to production. The discount structure doesn't automatically adjust, and the overage rates are rarely favorable.

The Open-Weight Escape Valve: Self-Hosted Inference

Not every step in an agentic workflow requires a frontier model. Tool-calling, classification, output validation, and routing decisions can often be handled by capable open-weight models like GLM-5 or Qwen 3.5 at near-zero marginal cost when self-hosted. Running a lightweight model for the 80% of agentic sub-tasks that don't require frontier reasoning, and reserving Claude or GPT-4o for the steps that genuinely need it, can materially reduce total token spend.

For teams that want the cost benefits of open-weight inference without the operational overhead of running their own GPU cluster, Together AI (scored 8.0/10 by the TopReviewed AI panel) offers managed inference for open-weight models with straightforward per-token pricing. The economics are significantly more favorable than frontier model APIs for high-volume sub-tasks.

For latency-sensitive agentic loops where the speed of each LLM call affects end-to-end task completion time, Groq (scored 7.7/10 by the TopReviewed AI panel) provides high-throughput inference on custom Language Processing Units. In agentic workflows where a slow reasoning step blocks downstream tool calls, inference speed has a direct cost implication: slower loops run longer, consume more wall-clock compute on surrounding infrastructure, and degrade user experience in ways that push teams toward over-provisioning.

The honest caveat: self-hosted inference shifts costs from token billing to infrastructure and engineering. The math only works above a meaningful usage floor. If your agentic workload is running a few hundred tasks per day, managed API pricing from frontier providers will likely be cheaper than the GPU time and engineering overhead required to run your own inference stack.

Four Mitigation Strategies That Actually Work in Production

Model Routing. Deploy a lightweight classifier, either a small open-weight model or a rules engine, to evaluate each incoming task and route it to the cheapest model capable of handling it. Frontier model calls should be reserved for tasks that demonstrably require frontier capability. Most agentic workflows contain a mix of complex reasoning steps and simple classification or formatting steps. Treating them identically in terms of model selection is where unnecessary spend accumulates.

Prompt and Semantic Caching. Anthropic's prompt caching feature allows repeated context blocks to be cached and re-used without re-billing the full token count on each call. For agentic workflows that re-inject the same system prompt, tool definitions, or retrieved documents across many loop iterations, caching can reduce effective token costs on those repeated segments. Redis-based semantic caching can extend this to near-identical sub-tasks across different workflow runs, not just within a single session.

Human-in-the-Loop Gates. Inserting mandatory human approval checkpoints before high-token, high-stakes agentic branches execute is a safety measure and a cost gate simultaneously. A runaway agentic loop that retries a failing sub-task 15 times before timing out doesn't just produce a bad outcome. It produces a very expensive bad outcome. Approval gates on branches that trigger deep reasoning chains or large context retrievals prevent the worst tail-cost scenarios.

Workflow Audit and Token Profiling. Instrument every LLM call in your orchestration layer with token counts before deploying to production. This is not optional. Most teams that do this for the first time discover that a small fraction of workflow steps, often around 20%, account for the large majority of token spend. That concentration makes optimization tractable. Without instrumentation, you're optimizing blind. Kestra (scored 7.4/10 by the TopReviewed AI panel) is an open-source workflow orchestration platform with observability hooks that make per-step token profiling tractable without building custom logging infrastructure from scratch.

What CFOs and Engineering Leaders Need to Align On Now

The organizational gap here is specific: engineering teams are building and deploying agentic workflows on timelines driven by product roadmaps, while finance teams are still budgeting on 2024 conversational AI assumptions. Those two things are on a collision course, and the collision happens at mid-year budget review, not at year-end planning.

Consumption-based AI billing requires the same financial controls that mature cloud infrastructure spending demands: budget alerts at the workflow level, per-workflow cost attribution that maps token spend back to a specific business process, and monthly cap reviews with hard stops before overages compound. The teams that built these controls for AWS and Azure spend in 2018–2020 need to apply the same discipline to agentic AI token costs now.

Anthropic's billing restructuring is a forcing function. Enterprises that haven't modeled agentic token consumption will not discover the gap at year-end planning. They'll discover it mid-year, when a production agentic workflow has been running at scale for three months and the invoice doesn't match anything in the approved budget.

Before any agentic workflow goes to production, run a 48-hour load test at real task volumes, instrument every LLM call with token counts, and then multiply the observed daily token rate by 3x as your budget baseline. Agentic usage scales faster than initial estimates in every production environment that has been observed publicly. Build that growth into the number before you sign off on deployment, not after the first billing cycle arrives.