The Agent Token Economy
The annual field guide to what AI agents actually burn: per-task token counts, cost per solved task and framework overhead across 30 production agents, 22 priced models and 3 public benchmarks.
Free 8-page executive summary. Full report $499 · Enterprise briefing $1,999.
- Where the tokens go: burn by domain, with each domain's share of the daily total.
- Cost per solved task across 30 production agents — from $0.011 to $35.71.
- Framework overhead measured against a raw API loop: +0% to +1,056%.
- Model price/performance: 22 models priced on the same 75/25 input/output workload.
Delivered by email as a PDF. No sales call, one-click unsubscribe.
Coding agents account for 89.9% of the tokens burned across the tracked fleet, yet the cheapest cost per solved task in the dataset is $0.011 — a 3,200× spread in what a "solved task" actually costs.
Recomputed from /data on every page load; the print edition freezes this at the edition date.
What is inside
Six chapters, 36 pages. Every chart is rebuilt from the same public dataset that powers this site, then frozen at the edition date.
-
Where the tokens go: burn by domainCoding agents account for nearly nine in ten tokens in the public dataset. This chapter maps the other five domains — data, research, ops, enterprise and marketing — with daily burn and cost per solved task for each.
-
Cost per solved task across 30 production agentsThe measured spread runs from about a cent to more than $35 per successful outcome. Every agent's retries, failed runs, model stack and success-rate denominator, normalised to one task definition.
-
Framework overhead: measured, not guessedMatched-task measurements put AutoGen at the baseline, LangGraph at +11%, the OpenAI Agents SDK at +25% and CrewAI at more than 11× the baseline. What the extra tokens buy, and when a heavier framework is still worth paying for.
-
Model price/perf: 22 models, one workloadEvery model priced on the same 75/25 input/output mix: from $0.20 to $20.00 per million blended tokens, with cache pricing, context windows and measured quality next to the price.
-
Retry and compaction wasteRetry overhead and failed-run burn across the dataset: where the second attempt costs more than the first, and which agents compact context instead of re-reading a million-token window on every turn.
-
Forecasts and the 2027 agent budgetThree scenarios for agent token spend through 2027, a per-team budget model, and the workload mixes that make a bill grow with solved tasks instead of headcount.
Straight from the dataset
A live preview of four charts from the paid report. The full edition adds per-agent tables, retry decomposition and the 2027 forecast.
Who it is for
The report is written for the people who approve, build or underwrite agent budgets — not for people shopping for a framework.
VCs & analysts
Underwrite the agent economy with unit economics instead of demos: cost per solved task, retry drag and where the next order of magnitude of token demand comes from.
- Comparable per-task cost across 30 production agents
- Where framework overhead silently eats margin
- A 2027 budget model you can put in a memo
Platform & infra teams
Benchmark your stack against real deployments, see whether a heavier framework is paying for itself, and build a token budget the finance team will accept.
- Framework overhead measured on matched tasks
- Model price/perf on one like-for-like workload
- Retry and compaction waste you can actually cut
Enterprise architects
Compare deployment shapes before committing a model contract: token footprint per task, context-window pressure and the orchestration overhead that shows up on the invoice.
- Per-domain burn and cost per solved task
- Model price spread with cache economics
- Briefing format for architecture review boards
How it is produced
AgentBurn keeps one open dataset of agent economics: 30 production agent profiles, 22 model price sheets and 3 public benchmark tables. The report is a frozen, edited cut of that dataset, extended with the analysis that does not fit on a leaderboard.
Measured first, modelled second
Every per-task number is either measured on a public benchmark, paper, vendor doc or telemetry stream; community-reported and clearly labelled; or modelled from official list prices. No figure appears without a provenance line, and estimates stay estimates in the tables.
The arithmetic is public
Cost per solved task is cost per task ÷ success rate. Token windows are runs per day × tokens per task × days. Framework overhead is matched-task overhead against the leanest baseline. The report uses the same formulas as the site, so any reader can reproduce a table from the JSON before trusting a conclusion.
Independent by construction. No vendor pays for a placement, a ranking or a favourable chapter. Sponsorship on the public site is labelled and never enters a computed metric — read the sponsorship policy.
Get the report
Start with the free summary, upgrade when the data is worth a budget line. Invoiced in USD, no subscription required.
Executive summary
The argument and the four headline charts, in a format you can forward to your team.
- 8-page PDF executive summary
- The token-burn map and cost-per-solved-task spread
- One data-refresh note when the edition updates
- No card, no call, unsubscribe in one click
Delivered by email. Add the full report later and the summary is credited against it.
The Agent Token Economy
The complete 2026 edition for one team. Every table reproducible from the shipped dataset.
- 36-page PDF with six chapters and full appendix
- Dataset export (JSON) frozen at the edition date
- Recorded 20-minute briefing walkthrough
- Quarterly data-refresh notes for the edition year
- Internal team licence — share it with your org
Stripe invoice or wire, net 15. Priced for one company, not per seat.
Enterprise briefing
The report plus a live session with the team that compiled it, scoped to your agents.
- Everything in the Full report
- Live 90-minute briefing with Q&A for up to 15 people
- One custom cut of the data for your stack or portfolio
- Follow-up memo with the answers we could not give live
- Invoice and security review support
Scheduled within 3 weeks. Refundable up to 48 hours before the session.
Report questions, answered
What exactly do I receive?
How often is it updated?
What is the refund policy?
How is my email data handled?
Is the dataset open if I buy the report?
Can I share it with my team?
Start with the free summary, then decide.
The four charts on this page are a preview. The report adds the per-agent detail, the retry decomposition and the budget model for 2027 — and the summary costs nothing but an email.