Elia Studio
Book a time slot

← All articles

Total cost of an AI solution: beyond tokens

Alexandre Gon · Head of AI & CTO4 estimated min read

Comparing AI solutions on token price leads to bad decisions. The seven cost components of TCO and how to price your exit before signing.

Cover: the total cost of an AI solution goes far beyond token price

Short answer: the total cost of an AI solution includes models, integration, data, security, evaluation, operations and exit. A price discount is not enough to choose between two offers. Compare cost per accepted result, contractual constraints and a realistic migration scenario before any durable production commitment.

Definition: total cost of ownership (TCO) of an AI solution. The sum, over the intended lifetime, of direct costs (models, infrastructure) and indirect costs (integration, security, evaluation, operations, exit).

What a 40% discount does not tell you

On 27 October 2026, Anthropic announced an expanded agreement with the GSA, the US federal procurement agency: a 40% discount on Claude Enterprise, Claude for Government and the API, extended to state and local governments, with an estimated saving of one billion dollars (official announcement, 27 October 2026).

Two details deserve more attention than the percentage. First, agencies are bound by no minimum spend commitment: they pay for what they consume. Second, the context is the US public sector, with specific security requirements (FedRAMP High) — this price does not transpose as-is to a European company, in a market where consumption already represents 80% of Anthropic's half-year revenue (CNBC, 27 October 2026).

The negotiation lesson travels well: the largest buyer got the discount and the freedom. When you compare two offers, both dimensions are on the table.

The seven components of total cost

  1. Models and calls. Tokens consumed, but also retries generated by unusable answers.
  2. Integration. Connectors, orchestration, testing, failure recovery.
  3. Data. Quality, preparation, usage rights, storage.
  4. Security and compliance. Access controls, logging, subprocessor terms.
  5. Evaluation. Test sets, regression measurements, production thresholds.
  6. Operations. Monitoring, observability, support, version upgrades.
  7. Exit. Migration, double run, training, loss of knowledge.

On a real project, components 2 to 7 typically outweigh component 1 from the second year onwards.

Pricing the exit before signing

Illustrative example for a document assistant used by 50 people:

  • extracting and rebuilding evaluation sets: 5 days;
  • rebuilding connectors: 15 days;
  • double run during the switch: 2 months at doubled cost over the period;
  • team training: 3 days.

Total: about twenty days of effort and two months of calendar time, excluding operational risk. Compare that amount with the annual saving the discount promises. If the saving is smaller than the exit cost, you did not get a discount: you signed a rent.

Managed API or private hosting

CriterionManaged APIHosted open-weight models
Entry costLowHigh (infrastructure, skills)
Cost at high volumeLinearDecreasing
Data locationPer contract and regionControlled
Skills requiredLow to mediumHigh
Vendor dependencyHighLow to medium
Upgrade paceVendor-drivenControlled

No column wins on every criterion. Sovereign architecture choices are made use case by use case. In production, the most frequent answer is mixed: route each task to the cheapest model tier that passes your quality thresholds, and track cost per accepted result.

Which option to choose

  • Managed API, if time to market dominates, the use case is reversible and the volume is low to medium.
  • Hosted open-weight models, if data must stay in your perimeter, or if predictable high volumes justify the investment.
  • A mixed architecture, in most real cases: per-task routing is the most common operating configuration we observe.

A poorly calibrated contractual commitment can wipe out the savings of a well-chosen architecture — the clauses to examine have their own article: AI contracts and consumption commitments.

Moving to the decision

Comparing AI solutions on list price is like choosing a car on fuel price alone. The sovereign AI roadmap and the FinOps workstream at Elia Studio systematically include this costing: cost per result, exit scenario, vendor dependencies. To apply it to your stack, book a call.

Frequently asked questions

Which costs should be added to token price?
Integration and connectors, data preparation and quality, security and logging, evaluation and test sets, operations and monitoring, human recovery from failures, and migration when the solution changes. These components often exceed the direct cost of model calls.
How do I compare a managed API with private hosting?
At equivalent output quality, compare cost per accepted result including operations. The managed API wins on time to market and low volumes; hosting open-weight models wins on predictable high volumes, data location and control over the lifecycle.
How do I price reversibility?
Estimate the cost of a full migration: extracting evaluation sets, rebuilding connectors, a double-run period, training. Compare that amount with the expected annual saving. If the saving is smaller than the exit cost, the discount never becomes an advantage.
Can a cheaper solution cost more?
Yes, in three ways: a cheaper model that generates more human rework; a proprietary architecture that raises migration costs; and hidden operations costs. Only cost per accepted result, measured on your real cases, settles the question.

↑ Back to top

Read next