Cost in the benchmark is a compound: the managed service, the selected model, tokenizer efficiency and the billing meter, measured together rather than one at a time [14]. Text is billed on input and output tokens, speech on audio duration, vision on the image or feature unit, so any workflow touching more than one modality is being compared across meters that do not convert into each other [7]. Even inside token-billed work, effective cost per outcome moves with tokenizer efficiency, prompt construction, context length and output size [8]. BCG's remedy is to price the business unit instead: cost per 1,000 summaries, per 10 hours of audio, per 5,000 image captions [9].
Three workloads across three providers gives nine provider-workload comparisons [22], and BCG's reading of them is that provider-level generalization carries a real risk of overpaying [13]. A rate won on the workload you happened to benchmark is a rate for that workload. The firm's framing is that AI economics are workload-specific rather than provider-specific [1], which is harder to put on a sourcing scorecard than a discount percentage. Of the three linked decisions, a headline token rate speaks to one [23], and that is the load-bearing part of BCG's warning that leaders who optimize on headline token prices will systematically overspend, under-govern, or both [3].
A CFO could fairly say this is an argument for paying someone to discover that prices are stable. BCG's Nimbus Pricing Index does report core cloud pricing holding steady, and BCG expects it to keep holding because hyperscalers need consistent cash flow to fund AI-related capital investment [5]. That stability is what drives the rest of the argument: with list prices flat, what remains variable is comparability, which BCG names as the buyer's real problem [6].
The board-deck version of all this is a slide showing unit cost down after a workload moved to the cheapest per-token option. What the slide cannot show is whether output held. In BCG's accepted samples, the gaps between the lowest-cost option and the others came from pricing structure, tokenization economics and service design, not from weaker outputs [18], which is why the firm pairs cost with a fit-for-purpose quality and latency rubric [19]. A unit-cost win becomes a saving only once a quality column sits beside it.
Sequencing is where this lands for anyone signing this quarter. BCG describes the operating model moving from individual assistants to teams of collaborating agents that distribute work across routing, retrieval and generation [21], and that distribution runs on orchestration, memory and guardrails chosen once and rarely reopened. A token rate settled this quarter is the smallest of the three decisions [23]; the architecture and the platform underneath it are being chosen in the same signature, with far less to check them against.
