Cost is one of the first questions raised about AI buying systems and one of the most difficult to answer well. Concourse Bench v1 (concourse.agency) provides the most detailed public dataset currently available on model API costs for a media-buying workflow. The figures are specific, bounded, and carefully qualified. The qualification matters as much as the numbers.

What did a completed buy cost in model API expense?

Among the 55 completed buys in v1, recorded model API expense ranged from $0.74 to $14.38 per completed buy. Elapsed time for completed runs ranged from 1.54 to 17.25 minutes. Both ranges are wide, reflecting variation between models, buying situations, and individual run behaviour.

The benchmark notes a definitional point worth taking seriously: cost per success is undefined for models with no completions. Luna spent $0.15 across all 12 of its attempts and completed nothing. Haiku spent $0.13 and also completed nothing. For deployment cost evaluation, what matters is the cost of a completed buy, not total API spend per model.

What did each model spend in total?

Known API costs across both briefs, as recorded in v1 (USD): Fable $21.04 (12 of 12 completed), Astra $12.32 (9 of 12 completed), Opus $12.20 (10 of 12 completed), Sonnet $7.42 (9 of 12 completed), Sol $5.85 (12 of 12 completed), Terra $3.15 (3 of 12 completed), Luna $0.15 (0 of 12 completed), Haiku $0.13 (0 of 12 completed).

Total known API cost for the full v1 suite was $390.16. The benchmark records up to $18.32 in additional charges that could not be resolved from the records, giving a total upper bound of approximately $408.48.

Does lower cost mean better value?

Not necessarily. The cost figures must be read alongside the completion rates. Fable had the highest total API spend at $21.04 and also the highest completion rate at 12 of 12. Its cost per completion was approximately $1.75. Sol spent $5.85 for 12 completions, a cost per completion of approximately $0.49. Astra spent $12.32 for 9 completions, a cost per completion of approximately $1.37.

Luna and Haiku had the lowest total API costs. They also completed nothing. The cost efficiency of a zero-completion model is not a useful metric for any deployment decision.

Among models that completed reliably, Sol produced the lowest cost per completed buy at approximately $0.49, against Fable at approximately $1.75. Opus sits at approximately $1.22 per completion, and Astra at approximately $1.37. These differences are real, but they are differences between models that all cleared the completion bar. The larger cost gap is between models that complete and models that do not.

Note: per-completion figures are derived from the benchmark totals and not published directly by Concourse Bench v1 (concourse.agency). Verify the arithmetic before relying on them for deployment decisions.

What do these figures include, and what do they exclude?

The benchmark is specific about the scope of the cost figures. They include recorded model API expense for the benchmark workload. They exclude: the cost of the buyer harness (the software that supplied the brief, checked every action, and recorded the deals), infrastructure costs, human oversight or review time, integration costs, and the cost of incomplete runs in a production context.

They are also explicitly not salary comparisons. Concourse Bench v1 states directly that the figures are "not full operating costs, salary comparisons or measured human savings." Any comparison of these figures to human buyer costs requires separate analysis of what a human completing the same workflow would cost, including time, overhead, and the cost of errors. No human baseline exists in v1; one is planned for a future version of the benchmark.

How should cost figures be used in deployment decisions?

The v1 cost data provides a useful input for scoping the API cost component of an AI buying deployment. It does not provide a complete cost model. A practitioner evaluating an AI buying system would need to add harness and infrastructure costs, oversight costs, the cost of incomplete runs, and the value of outputs beyond completion (buying quality, contractual protection, time savings) to produce a defensible business case.

The API cost range for a completed buy, $0.74 to $14.38, sits below the value of most programmatic media deals. But API cost is one component of total operating cost, and treating it as the whole figure would produce a misleading analysis.

There is a second consideration the data surface but do not resolve: the cost of failure. In v1, 41 of 96 registered attempts did not produce a completed buy. In a production deployment, incomplete runs still consume model API time, and the cost of an unfinished workflow includes not just the API expense but the time cost of human follow-up, the risk of a missed deadline, and the cost of the inventory opportunity not taken.

What comes next on cost measurement?

Concourse v2 will provide additional cost data under a stricter buying-quality test, with different tasks and rules. A human baseline is planned, which would allow cost per completion to be compared against the cost of a human completing the same task under the same conditions. Until that baseline exists, the v1 cost figures are informative about the API expense dimension of AI buying. They cannot support conclusions about total cost advantage or return on deployment.

All figures come from Concourse Bench v1, published by Alkimi at concourse.agency. The benchmark represents controlled simulation; figures should not be applied directly to live-market cost modelling without adjustment for the production environment.