Token Economics: How AI Agent Costs Really Arise

Token economics for AI agents describes the cost mechanics whereby every agent run is billed by the tokens consumed: input, output, cached and reasoning tokens. Unlike a chatbot, agents multiply consumption through multi-step loops, tool calls and sub-agents. A calculation based on the token price per request therefore underestimates real production costs many times over. Anyone planning an agent budget has to understand these mechanics, otherwise most of the bill is missing from the calculation.

• Tokens are the billing unit, not requests. As of 2026, a user request to an agent typically generates 5 to 20 LLM calls, each of which costs again.
• Output and reasoning are expensive, cached input is cheap. Output costs five times as much as input on Claude and six times as much on GPT-5.5; reused context (cache) costs at most 10 percent at Anthropic.
• API tokens are rarely more than 30 to 50 percent of total cost. Vector store, observability, compute and compliance drive the rest.

The four token types and how they are billed

Every LLM call breaks down into differently priced token buckets. Anyone who only sees the output systematically underestimates the input, because the agent drags along the system prompt, tool definitions and growing context again at every step.

• Input tokens comprise everything that goes into the model: system prompt, tool definitions, retrieved RAG context, conversation history and the actual request. A workflow with five defined tools at 150 tokens each adds 750 input tokens to every request on its own.
• Output tokens are the generated response. According to Anthropic's price list, they cost five times as much as input on all Claude models, and six times as much on GPT-5.5. Claude Sonnet 4.6, for example, stands at USD 3 input against USD 15 output per million tokens.
• Cached input tokens are reused, stable context. According to its price list, Anthropic reads cache at 0.1x the base price (90 percent discount), at 0.05x on Opus 5.5 and Sonnet 5.5, and at 0.025x on Fable 5.1 and Mythos 5.1. OpenAI charges around 10 percent of the base price on the GPT-5.x family. This is the most impactful cost lever introduced since the start of usage-based billing.
• Reasoning tokens arise from the internal thinking of reasoning models. As of 2026, they are simply billed at the output rate by both OpenAI and Anthropic. The practical consequence: a call with 20,000 reasoning tokens costs around USD 0.50 for the thinking alone at USD 25/million output, before a single visible word is generated.

The multiplier: why agents make consumption explode

The decisive difference from a chatbot is not the token price, but the number of calls. A single user request, which in 2023 was still one model call, routinely translates in 2026 into a chain: planner, tool selection, interpretation of the tool result, next-step decision, output formatting, often with explicit verification loops. That is 5 to 20 calls. Sub-agent cascades occasionally push this to 50 and more.

Three multipliers stack up:

• Multi-step execution multiplies consumption by a factor of 5 to 20 compared with a single call, depending on the number of steps. Each tool call is its own completion call with its own context.
• Sub-agent fan-out multiplies token consumption by a factor of 3 to 10 compared with a single agent. Each sub-agent is a separate completion with its own context window and its own tool definitions.
• Failure and retry add a factor of 1.3 to 3 with weak verification. Agentic workflows that fail and restart burn tokens on the way there.

On top of this comes context growth: with each step the conversation history grows, which is paid for again as input. Long contexts are doubly expensive: above 200,000 tokens, several providers (Gemini Pro models, OpenAI GPT-5.5 from 272,000 tokens) charge a surcharge of 2x input and 1.5x output. Naively stuffing the context window full is therefore rarely the cheap solution.

Hidden costs: the bill below the API line

For an agentic load at enterprise scale, the API token line is rarely more than 30 to 50 percent of total cost (TCO). Budgets that only apply the model provider's price list systematically miss at least half the bill. The following items are the most common blind spots in the DACH region (shares as a guideline for a representative enterprise load, as of 2026):

The hidden items in numbers: vector store and embeddings amount to 5 to 15 percent of total cost, observability to 2 to 8 percent, compute and sandbox to 10 to 25 percent. In the DACH region, factors are added that US budgets do not face: the EU-region surcharge of around 10 percent, the sovereignty premium of a factor of 1.5 to 3, as well as ongoing compliance costs of a realistic EUR 5,000 to 20,000 per year and active contracting partner. These DACH-specific factors increase total cost by 15 to 35 percent compared with a comparable US load.

The model-choice lever: small, large, or routed

Not every step needs the most expensive model. The second-biggest FinOps lever after caching is routing: the cheap model as the default, the expensive one only on a measurable gap. The price spread is considerable (as of 2026): according to Anthropic's price list, Claude Haiku 4.5 stands at USD 1 input / USD 5 output per million, Sonnet 4.6 at 3 / 15, Opus 4.6 at 5 / 25; according to OpenAI, GPT-5.5 stands at 5 / 30. At the low end, Claude Haiku 5.5 costs USD 0.10 / 0.50 for prompts up to 100,000 tokens, one fiftieth of Opus 4.6. DeepSeek charges USD 0.15 input for V4.1 Flash according to its price list, USD 0.30 at peak times, but via an API hosted in China.

As of 2026, Anthropic has formalized the routing pattern with the Advisor Tool (in beta since April 9, 2026): Sonnet or Haiku as the executor, Opus as an advisor brought in on demand within a single API call. The published benchmarks show how powerful the lever is: Sonnet plus Opus advisor reached 74.8 percent on SWE-bench Multilingual, Sonnet alone 72.1 percent. Even so, cost per agentic task fell by 11.9 percent compared with Sonnet alone. Haiku plus Opus advisor doubled the BrowseComp score (19.7 to 41.2 percent) at 85 percent lower cost than Sonnet solo.

The sober rule of thumb behind it: the cheapest model that passes the eval is the right model. At list price, Opus 4.6 (USD 5 / 25) costs roughly 1.7 times as much as Sonnet 4.6 (USD 3 / 15); in the 5.5 generation (Opus 5.5 at USD 4 / 20, Sonnet 5.5 at USD 2 / 10) the factor is 2. Whether that premium pays off only shows in the eval on the specific workflow. Teams that route by eval result rather than by gut feeling typically cut model costs by 30 to 60 percent with no loss of quality.

Worked example: 1,000 agent runs

Concretely, with Sonnet 4.6 rates (USD 3 input / USD 15 output per million, as of 2026, volatile). Assume a research agent with an average of 8 LLM calls per run, at 4,000 input and 800 output tokens each.

Unoptimized, without caching:

• Input: 1,000 runs x 8 calls x 4,000 tokens = 32 million tokens x USD 3 = USD 96
• Output: 1,000 x 8 x 800 = 6.4 million tokens x USD 15 = USD 96
• Direct token costs: around USD 192 per 1,000 runs

With an 80 percent cache hit rate on the entire prompt prefix (system prompt, tool definitions and previous history): according to Anthropic's price list, a cache hit on Sonnet 4.6 costs USD 0.30 per million instead of USD 3, one tenth of the price. For the remaining 20 percent you pay the base price or, if those tokens are newly written to the cache, 1.25 times the base price (5-minute cache). Input costs thus fall from USD 96 to about USD 27 to 32, the output remains. Total: around USD 123 to 128, a saving of around one third through caching alone.

If 60 percent of the calls are then routed to Haiku 4.5 (USD 1 / 5) instead of Sonnet 4.6, because the eval allows it, the direct line falls further toward USD 70 to 80 per 1,000 runs. And that is only the API line. If you add the vector store, observability and compliance ops, the real total cost is again significantly higher. This is precisely why per-workflow token caps (max_iterations, max_tool_calls, max_sub_agent_depth) are the governance standard as of 2026: they prevent a single run that has spun out of control from blowing up the calculation.

FinOps and budget guardrails

The effective measures are not secret knowledge, but engineering practice. Stacked together, a well-instrumented FinOps program delivers a 60 to 80 percent cost reduction compared with the unoptimized baseline:

• Aggressive prompt caching as the single biggest lever: a 60 to 90 percent cache hit rate on the prompt prefix cuts input costs by around 50 to 80 percent, because a cache hit costs one tenth of the base price on most Claude models. The 5-minute cache pays off after the first read access, the 1-hour cache after the second.
• Eval-driven routing via LiteLLM, OpenRouter or Portkey: the cheap model as the default, advisor/escalation patterns for the hard cases.
• Batch API for non-real-time loads with a flat 50 percent discount, combinable with caching. A cached batch request can thus fall to 5 percent of the standard price.
• Token budget per workflow with hard caps and cost attribution per tenant, team or workflow. Without this attribution, FinOps cannot answer the CFO's only question: which business unit is causing this bill?
• Open-weight fallback for the long-tail loads (summarization, classification, simple extraction): in the DACH region via EU-hosted providers such as Together AI EU region or DeepInfra Frankfurt, since the China-hosted DeepSeek direct API is ruled out for GDPR-bound loads.

For agencies and B2B decision-makers

Anyone running AI agents in production in the DACH region or building them for clients should make token economics a core competency, because the biggest cost levers lie not in the contract, but in the technical implementation. Procurement teams that wrestle over a five percent volume discount leave fifty percent on the table elsewhere. For agencies this means: cost attribution per client via Helicone or Portkey, transparent cost pass-through with a clear margin on the operational complexity (10 contracting partners mean 10 DPA chains), and sovereign hosting as a premium tier for clients with GDPR obligations.

For B2B the rule is: eval-driven model choice, caching from day one, per-workflow token caps and an exit path to open-weight providers for any load that exceeds a sensible monthly threshold. Blck Alpaca from Vienna supports DACH companies with precisely this calculation: from workflow architecture through FinOps guardrails to sovereign deployment. Note: all price figures in this article are as of 2026 and volatile; the price bands shift quarterly and should be checked against the current provider documentation before any budget decision.