How much did each version cost in API terms?

At API list prices, the four-way split came out to about $0.57 (range $0.41–$2.03) per article, versus about $0.42 (range $0.30–$1.49) for a single Opus agent, excluding the orchestrator's own tokens.

What were the two factual errors left in the split version?

One said there was no list of supported SoCs, when the official documentation does include one; the other carried an outdated maintenance deadline for Zephyr 3.7 LTS. Both were fixed after a second Opus re-check.

Was that checking step itself part of the difference?

Yes — the split version included verification by a separate Opus while the single-agent version only self-checked. Even excluding that verification, the split used 127,231 tokens, about 1.7 times the single-agent run.

Large Language ModelsAI Coding AssistantsZenn AI/ML10h agoNov 14, 2026, 16:00 JST

Claude Code sub-agent split costs 2.7× tokens of one Opus

Send on LINE

3 Key Points

What happened
On the same Canonical Zephyr article, the four-way Haiku/Sonnet/Opus split used 198,658 tokens over about 207 seconds, versus 74,381 tokens and about 168 seconds for one Opus agent.
Why it matters
The split version also left two factual errors after checking, while the single-agent version had zero, so extra agents bought more tokens and time but not more accuracy.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Get these free every morning

Context & Analysis

The test is a rare side-by-side of two ways to run Claude Code: one model doing the whole job, versus a team of role-limited sub-agents. The author has published token, time and error counts for both, but notes it is a single comparison on one topic — a Canonical explainer on long-term maintenance of the Zephyr embedded OS.

The mechanism the author offers is overhead. Each sub-agent starts with a blank state and reloads the project's CLAUDE.md and procedure files, and the fact-checker fetches the same web pages again that the collector already read. The largest single line item in the split run was Sonnet's writing at about 57,000 tokens.

What makes the result notable is that the split also lost on accuracy. The author's theory is that a model reviewing its own text tends to overclaim, so the revised setup keeps authoring and fact-checking in separate Opus agents. The author has not yet measured tokens for that configuration, so whether it recovers the savings remains untested.

FAQ

How much did each version cost in API terms?: At API list prices, the four-way split came out to about $0.57 (range $0.41–$2.03) per article, versus about $0.42 (range $0.30–$1.49) for a single Opus agent, excluding the orchestrator's own tokens.
What were the two factual errors left in the split version?: One said there was no list of supported SoCs, when the official documentation does include one; the other carried an outdated maintenance deadline for Zephyr 3.7 LTS. Both were fixed after a second Opus re-check.
Was that checking step itself part of the difference?: Yes — the split version included verification by a separate Opus while the single-agent version only self-checked. Even excluding that verification, the split used 127,231 tokens, about 1.7 times the single-agent run.

Zenn AI/MLRead Original Article

Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Get it free

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.