What Orchestrated AI Coding Agent Teams Actually Cost in Tokens
A token and dollar breakdown of five orchestrated AI coding agent pairings shows costs ranging from roughly $57 to $124 per task set.
What does it cost to run an orchestrated AI coding agent team?
Running a lead-and-worker AI coding team through six real tasks generated hundreds of millions of tokens and recorded costs ranging from about $57 to $124 per pairing, depending on which models led and which did the work. Across five tested pairings, recorded usage totaled roughly 936 million tokens and about $405 in orchestration spend, but incomplete logging means these numbers mark a floor, not a final bill.
TL;DR
- Token counts varied by more than 5x across pairings, with Opus leading Sonnet recording about 444.81 million tokens compared to 143.73 million for Opus leading Soul.
- Recorded dollar costs ranged from roughly $57 to $124 per pairing across the test set, with Opus-Sonnet the most expensive and Opus-Soul the cheapest among the two highest-scoring teams.
- Opus carries a steep unit price premium, costing $4.20 per million fresh input tokens and $20 per million output tokens, roughly double Soul and Sonnet’s $2 and $10 rates.
- Astra undercuts everyone on paper, at $10 per million fresh input and $0.50 per million cached reads, but its teams also produced the weakest scores on several tasks, including incomplete emulator and keyboard plans.
- Only 8 of 30 task runs had complete captured usage data, so none of the recorded subtotals can reliably crown a cheapest setup or a best score-per-dollar winner.
- Quota interruptions and provider errors repeatedly cut runs short, which both suppressed scores and likely undercounted true token usage for the affected pairings.
One coffee. One working app.
You bring the idea. Remy manages the project.
How was the token and cost data collected?
The test ran five lead-and-worker pairings through six coding tasks: an emulator, a Docker alternative, a room designer, a keyboard configurator, a Pokémon-style game opening, and a Fitbit desktop app. Each task was worth 10 points, for a possible 60 points per team. Every pairing got one high-effort run per task on a shared machine, routed through Claude Code, Codex, and a native CLI depending on the model, with up to three simultaneous workers, 20 lead planning rounds, and 50 assignments allowed per task.
The five pairings tested were Opus leading Soul, Astra leading Sonnet, Opus leading Sonnet, Astra leading Soul, and Opus leading a fourth model referred to as Muse Full. The lead model in each pair planned the work, assigned tasks to workers, and reviewed results. Token totals combined lead, worker, correction, and any additional recorded model calls, excluding setup and evaluation overhead. Resumed sessions counted usage only once, and any missing records stayed unknown rather than estimated.
That caveat matters: the teams did not run on equal computation budgets, there were no solo-model baselines, and roles were never reversed. That means the comparison can’t cleanly separate how much the orchestration itself contributed versus the raw capability of the underlying models.
How much did each pairing actually cost?
Using the saved API rates for each model, fresh input, cached reads, and output tokens were each priced differently. Soul and Sonnet each cost $2 per million fresh input tokens, $0.10 per million cached reads, and $10 per million output tokens. Opus cost more across the board, $4.20 per million fresh input and $20 per million output, making it roughly twice the unit price of Soul or Sonnet. Astra priced out at $10 per million fresh input, $1 per million cached reads, and $0.50 per million output, which works out to five times Soul and Sonnet’s fresh input and output rates and ten times their cached read rate.
For the fourth model, Muse Full, no confirmed tariff or subscription invoice existed, so the calculation used a standard assumption of $4 per million fresh input tokens, $0.20 per million cached reads, and $20 per million output tokens. That figure is explicitly not a confirmed rate.
Applying those rates to recorded usage produced the following subtotals: Opus leading Soul cost about $56.97, Astra leading Sonnet cost about $68.75, Opus leading Muse Full cost about $80.22, Astra leading Soul cost about $75.22, and Opus leading Sonnet cost the most at $124.22. Combined, that’s roughly $405.49 in recorded orchestration spend across all five pairings.
Why can’t these numbers crown a cheapest team?
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
Only 8 of 30 total task runs had complete, captured usage data. The rest are partial, missing, or affected by provider quota errors that cut workers off mid-task. Several runs, including Astra leading Soul on the Pokémon task and Opus leading Muse Full on the same task, were interrupted badly enough that they scored zero not because the model failed outright, but because quota or rate-limit errors blocked workers before they could deliver anything playable.
That means a lower recorded dollar figure for one pairing might just reflect missing data rather than genuine efficiency. The 936.11 million total recorded tokens broke down into about 926.74 million input tokens and 9.37 million output tokens, and most of that input was repeated cached context rather than fresh reasoning, which keeps the effective cost lower than the raw token count might suggest. Still, without complete logs across every task, none of these subtotals can fairly identify a cheapest setup or a best score-per-dollar winner.
Which pairing delivered the most for its cost?
Opus leading Soul finished with the highest overall score, 44.2 out of 60 (73.67%), while staying useful across all six tasks without major quota interruptions. Opus leading Sonnet followed closely at 43.7 (72.83%), but that run suffered quota problems across five of its six tasks, meaning its real capability may be understated by the recorded score and its $124.22 cost may not reflect a complete picture either.
Astra leading Soul trailed well behind at 26.3 (43.83%), partly dragged down by a completely empty, zero-scoring Pokémon run after an HTTP 429 error exhausted retries. Astra leading Sonnet performed unevenly: it produced the single highest score on two individual tasks (room designer at 9.3 and the Pokémon opening at 9.1) by catching design flaws other teams missed, like a walkthrough telling players to heal at a nonexistent healing center. But the same pairing scored just 1.9 and 2.8 on the emulator and keyboard tasks respectively, because its upfront plan never assigned the core execution work needed.
Judged purely on recorded cost against delivered score, Opus leading Soul looks like the strongest starting point: high performance, full six-task completion, and the lowest cost among the top two scorers.
Is cost or score the better signal here?
Neither tells the whole story alone. Cost figures are undermined by incomplete logging, so a cheap-looking pairing may simply have missing records rather than genuine efficiency. Score figures are undermined by quota interruptions that cut some teams off before they could demonstrate real capability. Astra leading Soul and Opus leading Muse Full in particular need complete, uninterrupted runs before any fair comparison on either axis is possible.
The more durable lesson is architectural: teams that assigned clear ownership over every required feature before work began (as Astra leading Sonnet did on the room designer, splitting model, editor, and renderer into separate owned pieces) produced more coherent results than teams whose initial plan quietly dropped entire execution paths, as happened with both Astra-led teams on the emulator task.
Frequently Asked Questions
Which AI model pairing was cheapest to run?
Among pairings with usable data, Opus leading Soul recorded the lowest cost at about $56.97, but only 8 of 30 task runs had complete usage logs, so no pairing can be confirmed as the overall cheapest.
Why does Opus cost more than Soul or Sonnet?
Other agents ship a demo. Remy ships an app.
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
Opus is priced at $4.20 per million fresh input tokens and $20 per million output tokens, roughly double the $2 and $10 rates for Soul and Sonnet, reflecting its position as a premium-tier model.
Does a cheaper model mean a better value?
Not necessarily. Astra priced far below Opus, Soul, and Sonnet, but Astra-led teams also produced some of the weakest scores, including incomplete plans on the emulator and keyboard tasks, showing that low unit cost doesn’t guarantee a usable result.
Why were so many task runs incomplete?
Several runs hit provider quota errors or rate limits mid-task, cutting off workers before they could finish. This affected Opus leading Sonnet across five tasks and caused complete zero scores for two other pairings on the Pokémon task.
Can these numbers be used to pick the most cost-effective AI coding team?
Not reliably yet. With only a fraction of task runs fully logged and multiple runs cut short by provider errors, the recorded totals establish rough cost ranges but can’t support a definitive cheapest-team or best-value ranking.
