What does fanning out 100 AI subagents cost? A calculation from measured runs

Key numbers

The charts

Hover or focus a row for its exact value, interval and sample. Each chart has a Table view.

Tables

Method

1. The calculation uses separate CLI processes and complete Crewshal worker attempts. It does not measure Claude Code’s built-in subagents within a session.
2. Minimal replies: 5 calls per CLI on one Mac, one at a time. Claude Code used Haiku 4.5; Codex used its default model. Wall time includes the answer and process exit. Claude input tokens include all cache categories.
3. The current public Codex aggregate reports 17,051 inclusive input tokens, including 13,184 cache-read tokens. Cached tokens are a subset; they must not be added again. We do not project a common startup cost across different models and prompts.
4. Small calls: 5 short tasks, 15 calls per selected configuration. We price the reported tokens at list rates, with recorded cache reads and writes. Failed validators remain in cost and time summaries. Pass rates use Wilson 95% intervals.
5. Worker costs: all 33 SWE-bench attempts, including failures, from 2 campaigns on 2 platform builds. We use recorded notional totals, including compaction, rather than reprice only the main-call tokens.
6. Cost scenarios multiply per-call medians or the mean worker cost by N. Serial duration scenarios multiply the median duration by N. Neither predicts a new workload.
7. First-turn cache shares (calculation from reported first-turn counters): Claude 18.7%, n = 6, range 18.7% to 18.7%; Codex 55.5%, n = 3, range 47.7% to 59.3%. These are process-level means, not pooled token shares.

Caveats

• Nothing ran in parallel. We did not measure rate limits, shared-host contention, machine limits or orchestration costs. One-unit medians are reference times, not parallel lower bounds.
• These are subscription calls priced at API list rates, not invoices. We did not measure subscription capacity or billing at scale.
• Minimal-reply calls do not isolate the CLI prefix. No empty-prompt control establishes tokens sent before work. Tools, prompts, flags and models differ across runs.
• Claude’s minimal-reply input median is 6,761; whole Sonnet calls have a median 2,081. Times also differ: 2.53 s versus 2.31 s. We did not test the cause. Each small call includes its own CLI prompt, not the separately measured minimal-reply prefix.
• Cache costs are scenarios, not bounds. Complete reuse assumes a cache write finishes before later calls start and their prefixes match. Actual writes can exceed plain-input prices.
• The 4 later Claude processes read fewer first-turn cache tokens than they wrote. Counters do not identify which ledger tokens they read. Claude used a shortened ledger; Codex used the earlier longer ledger.
• Pre-call protocol declaration is unverified from the selected public extracts. File times and final protocol text do not establish what was declared before the calls. This calculation supplies no new experiment or advance-declaration proof.
• The linked minimal-reply, head-to-head and cache extracts retain final protocol text with advance-declaration headings. No preserved pre-call protocol versions were supplied for these extracts. Their public summary labels this missing provenance before publication; historical raw records and review verdicts stay unchanged.
• The small task set is easy and selected, not a representative coding sample. Sonnet has three exact-text format misses with correct final numbers. Other selected configurations passed every call and hit a ceiling. Small n and overlapping cost ranges do not support a ranking.
• Other work shared the Mac and subscription accounts. Timing may include shared load; different dates, models and harnesses prevent causal comparisons.
• The worker sample has 33 attempts. Costs range from $0.89 to $4.83 and times from 1.6 min to 54.1 min. Campaigns used different platform builds and include replacement instances. New tasks may require different tokens and outcomes.
• These publication-label corrections need a fresh source and text recheck before publication. They do not establish advance declaration or replace the historical failed source review.

Sources

• Fan-out cost and serial time scenarios (calculation)

  Calculation · October 6, 2026

  N times recorded per-call median or worker mean cost. Serial duration uses N times median. Cache scenarios are assumptions, not bounds; one-unit medians are references, not parallel lower bounds. No parallel workload or capacity was measured.

  Raw data: routing-overhead/results.json, caching-consistency/caching.json, provider-h2h/receipts.json, swebench/attempts.json

• Routing overhead runs: policy microbenchmark and CLI start-up

  Our recorded runs · October 6, 2026

  In-process timing of the deterministic routing policy (20,000 timed decisions), CLI start-up with a one-word prompt (5 runs per CLI), and decision counts read from recorded bench runs. LLM router timings are reused from the routing runs.

  Raw data: routing-overhead/results.json

• Caching sessions and repeated prompts (Claude Code and Codex CLI)

  Our recorded runs · October 6, 2026

  Part 1: 5-turn CLI sessions over a fixed synthetic ledger, with the cache counters each provider reports per turn. Part 2: three prompts with deterministic validators, 10 repetitions per model. Declared protocols, validator controls before inference, every attempt kept; answers are published as ordinal ids, never as text.

  Raw data: caching-consistency/caching.json, caching-consistency/consistency.json

• Provider head-to-head: Claude Code models vs Codex efforts

  Our recorded runs · October 5, 2026

  Five short tasks with deterministic validators, declared protocol, every attempt kept.

  Raw data: provider-h2h/receipts.json

• Crewshal on SWE-bench Verified, campaign 1 (25 instances)

  Our recorded runs · October 4, 2026

  Stratified sample of 25 Verified instances (seed 20261004), one attempt each, official grading harness. Fixed model claude-sonnet-5-5, platform build f0ac3a8a.

  Raw data: swebench/attempts.json, swebench/exclusions.json

• Crewshal on SWE-bench Verified, campaign 2 (8 compiled-extension instances)

  Our recorded runs · October 5, 2026

  The 6 compiled-extension instances that campaign 1 could not run, plus 2 replacement candidates. One attempt each, platform build 236c0d3f.

  Raw data: swebench/attempts.json

• Anthropic list prices (Claude models)

  Vendor price list · September 21, 2026

  Prices as listed by the vendor on 2026-09-21 and recorded in the product price table. Cache reads at the listed rate, one-hour cache writes at twice the input price.

• OpenAI list prices

  Vendor price list · October 3, 2026

  Token prices as listed by the vendor on 2026-10-03.

Download the data

The study as JSON (with its sources), every chart point as CSV, or the whole public dataset. Free to reuse under CC BY 4.0: credit Crewshal and link the study.

Cite as: Crewshal public benchmarks, “What does fanning out 100 AI subagents cost? A calculation from measured runs”, updated October 6, 2026, https://agent.sasid.ai/benchmarks/subagent-fan-out-cost.

Write-ups on this study