An analysis of 23,747 billed turns across 34 days of Claude Code usage finds that subagent turns cost $0.082 on average versus $0.250 for main-loop turns, but the discount disappears once context size and output length are held constant — a subagent turn then costs the same as a main-loop turn. The apparent savings come purely from subagents carrying 3.5x less context and producing 13x less output, since they start with no inherited conversation history. The widely cited 'subagent tax' (4-6x cost multiplier) is real but describes a different scenario: fanning out to multiple agents that each reload a large shared context, multiplying cost instead of dividing it. A breakeven table shows delegation only pays off below 100k tokens of context for tasks of six or more turns, but pays off immediately above 250k tokens. Practical takeaways: delegate late in a session, keep subagent briefs and reports short, and only fan out for latency/independence, not cost savings, since parallel fan-out with large shared context is the expensive case.
Table of contents
What the split actually looks likeIs a subagent turn actually cheaper?Then why was my subagent bill only 4.6%?What is the "subagent tax" measuring, then?When does delegating actually pay?What this means for how you delegateHow I measured thisFAQThe part I did not expectQuestions this post answers
Are Claude Code subagents cheaper than doing the work inline?
Per turn, yes on average, but not because subagents are inherently cheaper. Across 23,747 billed turns, a subagent turn cost $0.082 versus $0.250 for a main-loop turn, but that gap comes entirely from subagents carrying 3.5x less context and emitting 13x less output. At equal context and output size, subagent and main-loop turns cost the same, within noise. Track real cost tradeoffs like this one on daily.dev as you tune agent delegation strategies.
What is the subagent tax people talk about when using AI coding agents?
It refers to the extra cost of fanning out a task to multiple agents that each load their own baseline context, measured elsewhere at 4.2x to 5.9x the tokens of doing the job directly. This tax is real when parallelizing across several agents, but it does not apply when delegating a single narrow task out of one large conversation, since that shrinks rather than multiplies the context carried. Developers weighing fan-out versus single-agent delegation can follow this cost analysis on daily.dev.
When does it actually pay to delegate a task to a subagent instead of continuing inline?
It depends on how large the main conversation already is: below about 100k tokens of context, delegation only pays off for tasks of six or more turns, while above 250k tokens it pays off from the very first turn. Spawning a subagent costs about $0.188 upfront because it must write roughly 34,000 tokens of cache before reading any back cheaply. Anyone tuning agent delegation thresholds can keep up with cost breakdowns like this via daily.dev.