Cohere

Foundation model platforms have shifted from a technical procurement question to a delivery economics question: the same client retainer now buys wildly different output depending on which model tier, context window, and retention terms an agency routes work through. Agencies that map deliverable types to model tiers and track token spend per account protect margin; those that default to one frontier model absorb cost variance they never bill for.

Model routing discipline treats model selection as a per-task decision rather than a per-agency default. Foundation model platforms price per million tokens and cap context per tier, so a retainer that runs every job through a frontier model burns margin on work a smaller tier handles at a fraction of the cost. The discipline has three moves: inventory the deliverable types an agency produces, assign each to the cheapest tier that clears the quality bar, then re-test quarterly as new releases land. OpenAI's October 2, 2026 guide splits the GPT-6 family into three variants for prototyping, feature development, and multi-step orchestration, which is exactly the tiering decision agencies now have to make explicit. Teams that route by task keep delivery costs predictable across client accounts; teams that route by habit pay frontier prices for draft work.

Retention Term Sheet treats data-retention language as a commercial constraint, not a legal footnote. Foundation model platforms differ on whether prompts and outputs are retained, for how long, and whether zero-retention or private deployment is available at a given tier. That single clause decides which client work an agency can route through a hosted API and which work must move to a private VPC or on-premises deployment. Cohere, for example, offers private VPC and on-premises deployment of Command, Embed, Rerank, Transcribe and Parse, which lets an agency accept a regulated client engagement that a shared-endpoint-only provider would block. The framework asks three questions before any build: what does the retention clause say, what does the client contract promise, and does the gap require a deployment change. Agencies that answer these at scoping time avoid rebuilding a delivery pipeline mid-retainer, and they can price the private-deployment premium into the original statement of work rather than absorbing it as margin loss.

Context Window Economics treats a model's maximum context length as a pricing variable rather than a feature checkbox. A 200K-token window lets an agency feed an entire client brand guide, three years of support tickets, and a product catalog into one call, which removes retrieval plumbing but multiplies input-token spend on every run. The framework asks two questions before any build: how many tokens does the average client deliverable actually consume, and does that volume justify a long-context model or a cheaper short-context model plus a retrieval layer? Prime Intellect's October 2026 launch of serverless and reserved serving for frontier open models gives agencies a third lever, since reserved capacity can cap the per-token rate on high-volume retainer work. Run the math per deliverable, not per token, and long-context defaults stop quietly eroding project margin.

Pick the model tier by mapping each recurring client deliverable to a tested tier, then verify the choice against your own eval set before it touches a retainer.

Contract a tested second provider and a documented cost-per-outcome model before you move any client retainer onto a single foundation model platform.

IF your agency ships client work through one model family and can absorb a provider outage without breaching a retainer, THEN concentrate spend and prompt engineering on that single platform to compound quality gains. IF delivery promises span multiple clients with different latency, context, and data-retention constraints, THEN route across at least two foundation model platforms and treat the router itself as billable infrastructure.