The Economics of AI Agents: Why Token Costs Are Falling and the Bill Is Still Going Up

Per-token prices for frontier AI models have dropped roughly 10x since GPT-4 launched in March 2023, and companies running agents in production are paying more, not less. The reason is mechanical. Agents send more tokens per task, run more tasks, and use models that think before they answer. Cheaper unit, larger basket. The bill goes up.

The price curve is steep and well documented

The unit price of intelligence is collapsing on a timescale measured in months.

GPT-4 at launch in March 2023 cost $30 per million input tokens and $60 per million output. Claude Sonnet 4.6 in May 2026 costs $3 per million input and $15 per million output, according to Anthropic's pricing page. That is a 10x drop on input, a 4x drop on output, a 6.7x drop on a blended 80/20 input-output basket. GPT-5, released August 2025, sits at $1.25 input and $10 output. Gemini 2.5 Pro is $1.25 input and $10 output for prompts under 200K tokens. Gemini 2.5 Flash is $0.30 input and $2.50 output, a hundredfold cheaper than GPT-4 launch input.

Andrew Ng put the GPT-4o number at $4 per million tokens blended, against $36 for GPT-4 launch, in August 2024. Epoch AI tracks the curve at roughly 10x per year on inference price for fixed capability, with the fastest cohorts dropping 40x and the headline cohorts dropping 900x annually since January 2024. The deflation is not a vendor anecdote. It is the dominant trend in the sector.

The implication everyone reaches for is wrong. Cheaper tokens do not produce cheaper bills.

What an agent actually spends

An agent is a multi-turn, tool-using process. It costs an order of magnitude more per outcome than a single prompt does, before any model upgrade.

A chat user sends one prompt and reads one answer. Maybe 4,000 input tokens, 1,000 output tokens. On Claude Sonnet 4.6 that runs $0.027 per interaction. An agent solving the same business problem will load a system prompt, tool definitions, prior turns, retrieved documents, then loop: read a tool result, decide, call another tool, read again. A modest ten-turn agent task carrying 50,000 tokens of context per turn (history accumulates, tool outputs accumulate, the trajectory grows) sends roughly 500,000 input tokens and 50,000 output tokens. That is $2.25 per task on the same Sonnet 4.6. Same model, 83x the spend.

Augment Code's loop analysis shows the same shape from a different angle. By turn 10 of a multi-turn conversation, cost per call sits at roughly 7x cost of turn 1, because each API request resends the full trajectory. Anthropic's own multi-agent research uses around 15x the tokens of a single chat. One developer tracked 42 agent runs against a FastAPI codebase and measured 70% token waste, mostly from reading too many files, retrying failed attempts, and absorbing verbose tool output.

The unit got cheaper. The unit count exploded.

Thinking tokens cost more per outcome

Reasoning models break the old mental model where output tokens are what the user sees. They are not.

Claude with extended thinking, OpenAI's o-series, GPT-5 reasoning mode, and Gemini's thinking variants all generate internal chain-of-thought tokens before emitting a visible answer. Those internal tokens bill at the standard output rate. A 500-token visible answer with 4,000 tokens of thinking behind it costs 9x what the bare answer would, because the model paid for 4,500 output tokens to give you 500. On GPT-5.5, output is $30 per million tokens. On Claude Opus 4.7, $25 per million. On Gemini 2.5 Pro, $10 per million. Thinking is billed identically.

The trade-off is genuine. Reasoning models post higher benchmark scores on hard tasks. The arithmetic just runs the wrong way for cheap tasks. A reasoning model burning two thousand thinking tokens to confirm an invoice category costs more than a non-reasoning model would have charged to do the whole job. Pointing reasoning models at every task is the most expensive default in production AI today.

Verbose, capable, and pointed at harder problems

Three things compound at once. Each one looks small. Together they swallow the price drop.

Agents are more verbose. Bigger context windows (1,000,000 tokens at Claude Sonnet 4.6 standard pricing, 1,050,000 at GPT-5.4, 1,000,000 at Gemini 2.5 Pro) invite implementers to dump entire codebases, document corpora, and meeting transcripts into the prompt. The window is the budget; budgets fill. Agents are more capable, so teams point them at harder problems with more iteration. Anthropic's research has documented agent runs that produce solutions to bugs unconstrained models would walk away from, at $5 to $8 per task in token spend. And inference-time compute has become a knob the model adjusts on its own. The system that decides to think longer when the question is harder is the system that decides to spend more on your behalf.

A16z's State of AI study on OpenRouter traffic found enterprise LLM spending more than doubled from 2024 to 2025, to roughly $8.4 billion. Anthropic's annual recurring revenue went from $9B to $44B across 2026, doubling about every six weeks (MindStudio analysis, May 2026). Token volume is growing faster than token price is falling. That is the whole paradox in one line.

What an SME deploying agents should know

The bill is not random. It is the product of token volume per task, task volume per month, and model choice. All three are levers.

Token volume per task is mostly trajectory management. Long-running agents accumulate history and tool output in their context, and most of it stops being useful after a few turns. In our implementations, the single biggest cost reduction is summarising or trimming the trajectory aggressively, keeping the model in a tight working window rather than the full 1M-token ceiling. The Stevens Institute Online piece on agent unit economics calls this the trajectory tax. Every team running multi-turn agents pays it.

Task volume is a business question, not an engineering one. An agent that runs once a day on demand is not the same cost profile as an agent embedded in a workflow that triggers a hundred times an hour. Scope the deployment to the value of the answer, not the capability of the model.

Model choice is the underused lever. Routing simple queries to Gemini 2.5 Flash ($0.30 input, $2.50 output) and reserving Claude Opus 4.7 or GPT-5.5 for the steps that genuinely need frontier reasoning can cut spend by 50 to 75% on most workflows without measurable quality loss. Research from the hierarchical agent literature shows budget executor models running under a frontier orchestrator can hit 97.7% of full-frontier accuracy at 61% of cost. Use the expensive model where it earns its price. Use the cheap model everywhere else.

What changes in 2026

The vendors have built the cost levers. Most teams have not turned them on yet.

Prompt caching is the largest gain available. Anthropic prices cache hits at 10% of standard input, with writes at 1.25x (5-minute tier) or 2x (1-hour tier). For an agent that resends the same system prompt and tool definitions every turn, an 80% cache hit ratio on a 500,000-token input drops input cost from $1.50 to $0.42 per task on Sonnet 4.6, a 72% reduction. The break-even is one cache hit. Most agent loops do not use it.

Batch APIs cut another 50% off both input and output for asynchronous workloads. The two stack: batch plus caching takes effective cost to roughly 5% of the headline price on eligible workflows. Anthropic, OpenAI, and Google all expose this. It is the cheapest infrastructure decision an implementer can make for non-interactive jobs. Model routing is the third lever and the one that requires actual engineering. Gemini 2.5 Flash, Claude Haiku 4.5 ($1 input, $5 output), and GPT-5.2 ($0.875 input, $7 output) are competent enough to handle most steps of most workflows. The frontier model is for the step that actually needs the frontier.

For an SME budget planning agentic deployment in 2026, the working number we use at Ellypsis is $50 to $500 per month per low-volume agent (a few dozen runs a day, well-engineered), and $1,000 to $5,000 per month per high-volume agent (hundreds of runs per hour, with caching and routing in place). Without caching and routing, multiply by three to five. The implementations that come in over budget are almost always the ones that did not turn the cost levers on.

The unresolved question

The honest part is that nobody knows where this lands.

Price per token is falling faster than tokens-per-task is rising, on the headline numbers. But tokens-per-task is rising faster than most companies' AI budgets are. Inference is projected to overtake training in hyperscale capex spend by 2026 (Gartner, March 2026 forecast), which means the cost curve for serving agents is still being built. Gartner also projects 90%+ cost reduction for trillion-parameter inference by 2030 against 2025 prices. Whether that arrives fast enough to absorb the demand from a workforce that, by then, is running thousands of agents per employee, nobody is willing to commit to in writing.

What we tell clients is the same thing every quarter. Treat the AI bill as a unit-economic problem, not a software licence. Measure cost per useful outcome, not cost per million tokens. The first number is the only one that survives the next price cut.

Token cost is one variable inside a larger implementation question. For how SMEs actually budget AI rollouts end-to-end, see how much an AI implementation really costs. For the connectivity standard that makes agent tool use possible, see what MCP is and why it matters for your business.