Evidence correction, 14 October 2026: earlier versions of this post attributed an agent token study to Microsoft Research and misreported its per-task token figures, gave the wrong period for the token price decline, presented an API aggregator's own platform data as industry data, cited survey and cost-breakdown figures their sources do not contain, and overstated an example team's savings about ninefold. Those claims are removed or corrected below.
The Jevons Paradox, applied to tokens
In 1865, William Stanley Jevons observed that making coal engines more efficient did not reduce coal consumption. It increased it. Cheaper energy per unit made new uses viable, and total demand outpaced the efficiency gains.
The same thing is happening to LLM tokens.
Stanford's 2025 AI Index found that the inference cost of a system performing at the level of GPT-3.5 dropped over 280-fold between November 2022 and October 2024, from about $20 to $0.07 per million tokens. Anthropic dropped Opus from $15/$75 per million tokens to $5/$25. OpenAI released GPT-4o mini at a fraction of GPT-4's cost. Google pushed Gemini Flash below $0.10 per million input tokens. DeepSeek and open-source models compressed the bottom end even further.
Enterprise spending rose anyway. Menlo Ventures estimates that companies spent $37 billion on generative AI in 2025, up from $11.5 billion in 2024, more than three times as much. Gartner forecasts worldwide AI spending of about $2.5 trillion in 2026. The FinOps Foundation's State of FinOps 2026 found that 98% of respondents now manage AI spend, up from 31% two years earlier.
Cheaper tokens did not reduce the bill. They changed what was possible. And what became possible consumes far more tokens per task.
What changed: the shift to agentic workloads
In 2024, most production LLM usage was single-turn: a user sends a prompt, the model responds, done. One API call per interaction, predictable token counts, manageable costs.
By 2026, the dominant pattern is agentic. A coding assistant reads a task, calls a tool, reads the output, decides what to do next, calls another tool, reads that output, retries if the test fails, and repeats. A single task that used to be one API call can now trigger dozens of calls, each carrying the accumulated context.
Researchers from the University of Michigan and Stanford quantified this in April 2026, analyzing runs by eight frontier models on SWE-bench Verified. They found:
- Agentic coding tasks consume about 1,000x more tokens than code reasoning and code chat tasks
- An average agentic task used about 4.2 million tokens, with roughly 154 input tokens for every output token
- The same agent on the same task can vary by up to 30x in total token consumption between runs
Stanford's Digital Economy Lab, whose researchers co-authored the paper, summarized the key finding: input tokens, not output tokens, drive the cost, because the agent re-reads the accumulated context before each action.
Gartner's March 2026 analysis put the multiplier at 5 to 30x more tokens per task for agentic models compared to a standard chatbot. That range depends on the task and the agent framework. Against a 10x price cut, the high end (30x the tokens at a tenth of the price) triples the bill, while the low end (5x the tokens at a tenth of the price) halves it.
The math that explains your bill
Here is an illustrative example with hypothetical numbers for one enterprise:
| Factor | 2024 | 2026 | Change |
|---|---|---|---|
| Average cost per million tokens | $15.00 | $5.00 | -67% |
| Average tokens per task (single-turn) | 2,000 | 2,000 | 0% |
| Average tokens per task (agentic) | N/A | 50,000 | New workload |
| Share of workload that is agentic | 5% | 45% | +40pp |
| Monthly API calls | 50,000 | 200,000 | +300% |
The per-token price dropped 67%. But the number of tokens per task grew 25x for nearly half the workload, and total call volume quadrupled as teams automated more with agents. The bill went up, not down.
This is the Jevons Paradox in action. Cheap tokens made it economically feasible to run agents on tasks that nobody would have automated at $15 per million tokens. Once the price crossed a threshold, usage exploded. The efficiency gain (cheaper tokens) was more than offset by the demand increase (more tokens per task, more tasks automated).
Where the money actually goes
NavyaAI, an AI cost-audit firm, published a cost report in February 2026 arguing that the model invoice is only part of the story. It claims 72% of production AI cost sits outside the model invoice, in orchestration, retrieval, retries, and observability.
Inside the model bill, the cost drivers for a team running agentic workloads look like this:
- Context accumulation. Every turn in an agentic session re-sends the conversation history, so later turns carry more context than early ones. In the April 2026 study's phase analysis, cache-read input tokens, the re-sent context, were the largest cost category in every phase of a run.
- Retry loops. When an agent produces code that fails a test, it retries. Each retry carries the accumulated context plus the error message. The same study found that the most expensive runs had sharply more repeated file views and edits.
- Tool schema overhead. Agentic frameworks inject tool definitions into every API call. A coding agent with many tools can add thousands of tokens of schema to every request, whether or not those tools are relevant to the current step.
- Productive reasoning. More tokens do not buy more accuracy. The study found that accuracy often peaked at intermediate cost and saturated at higher costs.
This is why "just negotiate a volume discount" does not fix the problem. The waste is structural, baked into how agentic workflows accumulate context. A volume discount lowers the price of each token, but it does not change how many tokens an agent re-sends.
Why cheaper models alone do not fix it
The intuitive response is to switch everything to the cheapest model. Run Haiku at $1 per million tokens instead of Opus at $5, and the 5x price difference should absorb the token growth.
In practice, this breaks on complex tasks. A coding agent using Haiku to architect a distributed system produces worse output, fails more tests, triggers more retries, and can end up costing more in wasted tokens than Opus would have. The retries compound: each failed attempt adds to the context, making subsequent attempts more expensive regardless of the model.
As an illustrative assumption, suppose an agent's turns split like this:
- 60 to 70% of turns are low complexity. File reads, status checks, formatting output, parsing errors. These are classification-grade tasks.
- 20 to 30% are medium complexity. Writing a function, generating a test, explaining a bug.
- 5 to 15% are genuinely hard. Architecture decisions, multi-file refactors, complex debugging.
Running everything on Opus means paying frontier rates for file reads. Running everything on Haiku means degrading quality on the hard turns that actually need reasoning power. Neither option addresses the structural problem.
The lever that works: per-request model routing
If 65% of the tokens in a session hit a model that is 5x more expensive than necessary, and you fix that, you cut the bill roughly in half (65% of the spend, 80% cheaper, is a 52% saving) without touching the complex tasks that need the expensive model.
This is what intelligent model routing does. A classifier evaluates each request (not each session, each individual API call) and sends it to the cheapest model that can handle that complexity level. The classifier adds some latency to each request, so measure it on your own traffic.
Here is an illustrative 40-turn agentic session, priced at input-token list rates only:
| Segment | Turns | Model (no routing) | Model (routed) | Cost (no routing) | Cost (routed) |
|---|---|---|---|---|---|
| File reads, status checks | 25 | Opus ($5/M) | Haiku ($1/M) | $2.25 | $0.45 |
| Code generation, tests | 10 | Opus ($5/M) | Sonnet ($3/M) | $1.40 | $0.84 |
| Architecture, debugging | 5 | Opus ($5/M) | Opus ($5/M) | $0.88 | $0.88 |
| Total | 40 | $4.53 | $2.17 |
That is 52% savings on this example. The five turns that genuinely need Opus still get Opus. The 25 turns that are reading files no longer pay frontier rates for it.
At $2.36 saved per session ($4.53 minus $2.17), a team running 200 sessions per week (about 867 sessions per month) would save roughly $2,000 per month, and about $5,100 per month at 500 sessions per week. These are hypothetical figures from the example above, not measured results.
The industry is already moving
AI.cc, an API aggregator, reported that its enterprise accounts called an average of 4.7 models in Q1 2026, up from 2.1 a year earlier.
AI.cc also reported that accounts using three-tier routing on its platform paid a median $2.31 per million tokens in 2026, against $18.40 for frontier-only workloads a year earlier. That gap mixes a year of price cuts with routing, and the data is the vendor's own.
IDC predicts that by 2028, 70% of top AI-driven enterprises will use multi-tool architectures to dynamically manage model routing. The question for most teams is not whether to adopt routing, but when.
What you can do this week
1. Measure your actual token distribution. Pull a week of API logs. Bucket each request by complexity (simple lookup vs. code generation vs. hard reasoning). See what share of your calls actually needs a frontier model.
2. Calculate your theoretical savings. Take the simple and medium requests from step 1 and price them at Haiku ($1/M) and Sonnet ($3/M) rates instead of Opus ($5/M). That gap is the most routing could save, before you check quality on the routed requests.
3. Start in shadow mode. Nadir's Decision API targets about 100 ms, calls no model provider, and records model, effort, cache, context, and policy evidence. NadirClaw is source-available under PolyForm Noncommercial and runs locally. Attach actual outcomes before enforcement. Nadir's only charge is a flat $0.002 per request, never a share of savings.
The Jevons Paradox is not going away. Tokens will keep getting cheaper, and that will keep enabling new use cases that consume more of them. The only durable strategy is to make sure each token hits the right model for the job.
Sources: [Stanford HAI, "The 2025 AI Index Report"](https://hai.stanford.edu/ai-index/2025-ai-index-report) (2025). [Epoch AI, "LLM inference prices have fallen rapidly but unequally across tasks"](https://epoch.ai/data-insights/llm-inference-price-trends) (March 2025). [Menlo Ventures, "2025: The State of Generative AI in the Enterprise"](https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/) (December 2025). [Gartner, "Gartner Says Worldwide AI Spending Will Total $2.5 Trillion in 2026"](https://www.gartner.com/en/newsroom/press-releases/2026-1-15-gartner-says-worldwide-ai-spending-will-total-2-point-5-trillion-dollars-in-2026) (January 2026). [FinOps Foundation, "State of FinOps 2026"](https://data.finops.org/) (2026). [Bai et al., "How Do AI Agents Spend Your Money? Analyzing and Predicting Token Consumption in Agentic Coding Tasks," arXiv:2604.22750](https://arxiv.org/abs/2604.22750) (April 2026). [Stanford Digital Economy Lab, "How are AI agents spending your tokens?"](https://digitaleconomy.stanford.edu/news/how-are-ai-agents-spending-your-tokens/) (May 2026). [Gartner, "Gartner Predicts That by 2030, Performing Inference on an LLM With 1 Trillion Parameters Will Cost GenAI Providers Over 90% Less Than in 2025"](https://www.gartner.com/en/newsroom/press-releases/2026-03-25-gartner-predicts-that-by-2030-performing-inference-on-an-llm-with-1-trillion-parameters-will-cost-genai-providers-over-90-percent-less-than-in-2025) (March 2026). [NavyaAI, "AI Cost Report 2026: Token Prices & Rising AI Bills"](https://www.navyaai.com/reports/ai-cost-report-token-prices-vs-ai-bill) (February 2026). [AI.cc press release, "AICC Report: Enterprise Token Costs Drop 67% Year-Over-Year as Multi-Model AI Adoption Hits Record High"](https://www.einpresswire.com/article/911544568/aicc-report-enterprise-token-costs-drop-67-year-over-year-as-multi-model-ai-adoption-hits-record-high) (May 2026). [IDC, "The future of AI is model routing"](https://idc.com/resource-center/blog/the-future-of-ai-is-model-routing) (November 2025). Anthropic, OpenAI, Google model pricing as of May 2026.