Article reader Listen + reading controls
Article reader
Preparing the reader…
Reading settings
Buying four frontier models is not yet a multi-model strategy¶
The Department of Defense's (DoD) Chief Digital and Artificial Intelligence Office (CDAO) details prototype awards to Anthropic, Google, OpenAI, and xAI on July 14. Each agreement carries a ceiling of $200 million and aims to develop agentic workflows for national-security missions.
The awards create access and competition. The next challenge is turning provider diversity into architectural leverage rather than four separate silos.
Artificial intelligence (AI) models differ in capability, latency, cost, deployment environment, security posture, and tool support. They also change quickly. A model that leads one evaluation this quarter may be replaced the next.
That makes a multi-model strategy attractive. It also makes governance more complex.
Portability has to be engineered¶
Using several vendors does not prevent lock-in if every workflow depends on a provider-specific interface, prompt pattern, evaluation suite, or data integration. True optionality requires common layers where they are useful:
- portable task and evaluation definitions;
- normalized logging and provenance;
- shared identity and policy enforcement;
- model-routing rules;
- data and artifact export;
- and contract terms that support transition.
Common does not mean lowest common denominator. Programs can use specialized features while preserving enough separation to compare and replace components.
Route by mission need¶
A defensible router should consider more than benchmark rank. A low-risk summarization task may favor speed and cost. A sensitive analysis may require a particular accredited environment. A coding workflow may need strong tool use and traceability. Some tasks may require independent outputs from different models to expose disagreement.
Routing rules should be explicit and testable. Otherwise, “multi-model” becomes manual vendor preference or an opaque optimization no mission owner can explain.
Build a common evidence plane¶
The Department can gain unusual insight by evaluating multiple frontier models against the same representative tasks and operational measures. The evidence plane should capture:
- task performance and error type;
- review effort and override rate;
- latency, cost, and availability;
- security and policy incidents;
- sensitivity to model updates;
- and mission outcomes after deployment.
That information can support acquisition and architecture decisions across programs without pretending one global ranking fits every mission.
Keep competition alive after award¶
Contracts should create pathways for new models and smaller suppliers to enter when evidence supports them. Data rights, interface documentation, and modular work statements can reduce the cost of change. Shared test harnesses can allow a challenger to demonstrate value without rebuilding the entire application.
The awards are a credible way to learn with multiple leading providers. They become a strategy when the Department can compare them on mission evidence, combine them deliberately, and change course without discarding the surrounding workflow.
Buying four models creates options. Architecture, evaluation, and acquisition discipline determine whether the options remain real.
Sources and research trail¶
- Chief Digital and Artificial Intelligence Office, “CDAO Announces Partnerships with Frontier AI Companies” (July 14, 2025).
- National Institute of Standards and Technology, AI Risk Management Framework 1.0 (2023).
- Office of Management and Budget, M-25-22 on federal AI acquisition (2025).
- Baldwin and Clark, Design Rules: The Power of Modularity (2000).