What does an AI agent really cost?
The $900 loop
On August 5, 2026, Pieter Levels (@levelsio) posted about running Matt Shumer's Gauntlet Loop, the viral one-prompt method that fans out subagents until each deliverable is "utterly wowed." His verdict was short: "a total mess and chaos of unperformant code." Each loop cost $500. Hours later he corrected himself: $900 total, and "I had to remove 95% of what it made and go back to what I had!"
Levels is not a beginner. He is one of the most followed solo builders in AI and has shipped more products than most agencies. If the most viral agent workflow of 2026 burned $900 a run on his stack, the pricing question changes for everyone: not "what does the model cost per token," but "what does the whole run cost, loops included?"
Why agent cost estimates hide the loops
Most estimates multiply list token prices by a projected request count and stop there. That math assumes every request does its job on the first pass. Real agents don't work that way:
- Subagent fan-out multiplies the bill. A small task that cost 121,000 tokens done directly cost 513,000 tokens fanned out to two subagents, a 4.2x jump. Each subagent re-reads its own context on every turn.
- Retry loops re-pay full context each pass. "Let me try something else" loops fail to converge, and infinite loops have racked up roughly $2,000 in API charges overnight, often enough that a budget-kill tool category now exists.
- Overhead bills before the work starts. Claude Code sends about 33,000 tokens of system prompt and tool schemas before your prompt arrives, and 75,000 to 85,000 with a real-world config.
- Effort settings scale cost directly. Max-effort mode on Claude Sonnet 5 uses up to 6x the agentic turns of low effort, and the measured cost is $2.29 per benchmark task.
- No budget rails. A developer's normal coding-agent workflow produced a $37,901.73 AWS bill from a prompt-caching miss, with no hard limit anywhere in the stack. His own framing: "This was not a leaked key. This was not crypto mining. This was not an infinite loop."
The surprise is the pattern, not the price level. One Cursor user's 70-day workflow averaged $0.06 per request, then produced a single $2.78 request and projected to about $416 a month. GitHub Copilot users on metered billing reported $6 on one feature request and 8% of a monthly quota burned in two hours; legacy multipliers price GPT-5.5 at 57x a base request. OpenClaw users hit $141 overnight from a misconfigured heartbeat, and $800 in a month on setups they considered normal.
The blowups, priced
| Scenario | Reported cost | What it shows |
|---|---|---|
| levelsio Gauntlet Loop | $500 per loop; $900 worst case; 95% of code removed | Fan-out + retries turn a viral method into a $900 burn |
| Subagent fan-out (two subagents) | 121,000 → 513,000 tokens (4.2x) | Each subagent re-reads its own context every turn |
| AWS Bedrock prompt-caching miss | $37,901.73 | No budget rails anywhere in the stack |
| Cursor 70-day workflow | $928.45 total; $0.06 average; $2.78 max per request | The tail, not the average, is the surprise |
| GitHub Copilot metered billing | $6 one feature request; 8% quota in 2 hours; GPT-5.5 57x (legacy) | Agentic work consumes far more compute |
| OpenClaw misconfigured heartbeat | $141 overnight; $800/month | Ambient background activity bills like full calls |
Budget the loops, not just the tokens
Three rules for anyone pricing or buying AI-assisted work:
- Ask what share of the quote is raw model usage versus human review and delivery. If the model bill is quoted high on models that cost cents per million tokens, that is margin, not cost.
- Model agent complexity in the estimate, not just tokens in and tokens out. Subagents, retries, context size, caching, effort level, and a budget cap all change the real number by an order of magnitude.
- Insist on budget rails. A kill switch and a hard cap turn a $37,000 surprise into a $50 one.
Agent cost is dominated by architecture and failure modes, not model price. Estimate the run the way agents actually run, and the estimate stops lying to you. When you are ready to compare, the agencies that price AI-assisted delivery transparently will show you the raw-model share, the complexity assumptions, and the budget rails — all three of the rules above, answered before you sign.
See how top AI agencies price AI-assisted delivery
Browse AI Agencies →Frequently asked questions
How much did levelsio's Gauntlet Loop really cost?
Pieter Levels (@levelsio) reported $500 per loop on August 5, 2026, then corrected the same day: $900 total, with 95% of the generated code removed in cleanup. Both figures are his own words.
Why do agent cost estimates hide the loops?
Most estimates multiply list token prices by a projected request count and assume every request succeeds on the first pass. Real agents fan out subagents, retry, reload full context each pass, and bill overhead before the work starts — which is where the cost explodes.
What is the biggest agent cost amplifier?
Subagent fan-out. A small task that cost 121,000 tokens done directly cost 513,000 tokens fanned out to two subagents — a 4.2x jump — because each subagent re-reads its own context on every turn.
How expensive can a single agent run get?
Public examples range from $0.06 average per request on a normal 70-day Cursor workflow, to $2.78 for one heavy request, to a $37,901.73 AWS bill from a prompt-caching miss with no budget rails anywhere in the stack.
How do I keep AI agent costs under control?
Budget the loops, not just the tokens: ask what share of a quote is raw model usage versus human review, model agent complexity (subagents, retries, context size, caching, effort level) in the estimate, and insist on budget rails — a kill switch and a hard cap.
How should I price AI-assisted work from an agency?
Ask the agency what share of the quote is raw model usage versus human review and delivery, how they model agent complexity and retries, and what budget rails they run. Agencies that price AI-assisted delivery transparently answer all three.
Sources
- levelsio X post, Aug 5, 2026 (incl. $900 / 95% correction): x.com/levelsio/status/2084997902632390981
- Matt Shumer, "How to Run a Gauntlet Loop": somethingbig.ai/gauntlet-loop
- Claude Code token overhead and subagent fan-out (121k to 513k): systima.ai/blog/claude-code-vs-opencode-token-overhead
- $37,901.73 AWS Bedrock bill from a prompt-caching miss: news.ycombinator.com/item?id=47933355
- 7 subagents burned a budget before one finished: news.ycombinator.com/item?id=48883275
- AgentGuard, budget-kill tooling for infinite loops (~$2,000 overnight): github.com/dipampaul17/AgentGuard
- Cursor 70-day usage, $928.45 total, $0.06 avg, $2.78 max: news.ycombinator.com/item?id=45931825
- GitHub Copilot metered-billing backlash ($6 one request, 8% quota in 2 hours): theregister.com
- Copilot model multipliers (GPT-5.5 = 57x, legacy): docs.github.com
- OpenClaw cost shock ($141 overnight, $800/month): clawsnewsletter.substack.com
- Claude Sonnet 5, $2.29 per benchmark task: artificialanalysis.ai
- "Let me try something else" retry loop: cimons.com