Agent economics · 2 November 2026
Claude Code vs Codex CLI: reading the cost tables correctly
24.1M tokens per task vs 3.9M. The gap is the harness, not the model — and it decides which table is relevant to you.
11 min readRead post →
Blog · agent economics
What AI agents actually cost
Working notes from the AgentBurn dataset: where agent tokens go, which metric survives procurement, how much of the bill is the framework, and what the published numbers leave out. Every figure is traceable to a project row, a benchmark sheet or a documented formula — and keeps its provenance label.
13 posts · last updated · RSS feed · by AgentBurn Editorial
Agent economics · 2 November 2026
How to estimate a monthly AI agent bill (with the formula)
Five inputs, four measurable, one a guess. Worked estimates for three agent shapes and the sensitivity that matters most.
10 min readRead post →
Agent economics · 2 November 2026
Retry overhead: the 31% of an agent bill nobody budgets
Debug-and-retry loops are the largest single token line in the Claude Code workflow split. How to measure the line and price it.
9 min readRead post →
Agent economics · 2 November 2026
Prompt caching and the agent bill: what a 90% hit rate is worth
Cached input is a fraction of fresh input — if the prefix is stable. A sensitivity model across the 22 priced models.
9 min readRead post →
Agent economics · 2 November 2026
Seats, credits, tokens: comparing AI agent pricing units
Agentforce sells 20 credits per action; Copilot Studio meters 2/5/10; coding agents quote per task. Normalize before comparing.
10 min readRead post →
Agent economics · 2 November 2026
Agent ROI: the break-even math against a human baseline
$1.17 per solved task is cheap against a $40 human task and expensive against a $2 one. Where the break-even line sits.
9 min readRead post →
Agent benchmarks · 2 November 2026
How to read a coding-agent leaderboard without being misled
SWE-bench Verified, Terminal-Bench 3.0 and the AA Coding Agent Index measure different things. A checklist for the rows.
10 min readRead post →
Measurement · 2 November 2026
When benchmarks stop helping: instrument your own agent
Benchmarks describe products, not your stack. What telemetry adds and the minimum viable measurement set.
9 min readRead post →
GEO and AI search · 2 November 2026
llms.txt and crawler policy for an agent data site
The exact machine-readable set AgentBurn publishes for AI crawlers, and the evidence for what each file does.
9 min readRead post →
Measurement · 2 November 2026
What 20 of 30 agent projects don't publish — and why that matters
Only three projects publish tokens per task. The provenance tiers that keep an agent cost database honest.
9 min readRead post →
Agent frameworks · 2 November 2026
Agent framework token overhead: AutoGen vs LangGraph vs CrewAI
379 tokens per run in AutoGen, 4,382 in CrewAI — an 11.6x orchestration tax measured on matched runs.
10 min readRead post →
Agent economics · 31 October 2026
Cost per solved task: the agent metric vendors leave out
An agent resolving 67% for $0.42 can be cheaper per solved task than one resolving 73% for $0.85. The arithmetic.
8 min readRead post →
Agent economics · 29 October 2026
AI agent cost per task: the unit that predicts your bill
The published spread runs from 0.137M to 24.1M tokens per coding task. The gap is the harness, the cache and the retries.
9 min readRead post →
What this blog is for
Most agent cost claims cannot be checked. A vendor says "up to 40% cheaper"; a blog post says "agents are expensive now". Neither gives you a unit you can reproduce. This blog does one thing: take a number from the open dataset or a cited benchmark, show the arithmetic, and state the limits.
Three rules apply to every post:
- Traceable numbers. Each figure either comes from a project row, a benchmark sheet or a formula on the methodology page — with the source link attached.
- No composite scores. Tokens per task, cost per task, cost per solved task and framework overhead stay separate. Averaging them hides the decision.
- Limits stated. Modelled run volumes, stale benchmark rows and list-price assumptions are labelled in the post, not buried in a footnote.