Enterprise AI should be organized around the verified business resolution, not the model used to produce it.
Frontier model access can now change because of government action, provider policy, price, or availability. That turns model dependence into a business-continuity and margin risk, not merely a procurement decision.
Frontier Model Access Is Now Conditional
On June 9, Anthropic launched Claude Fable 5, its most capable generally available model, and Claude Mythos 5, a less-restricted version initially available to selected cybersecurity partners. Three days later, the US government issued an export-control directive requiring Anthropic to suspend access to both models by any foreign national, including Anthropic employees who were not US citizens.
Because Anthropic had no reliable way to verify nationality in real time, it disabled both models for all customers. The controls were lifted June 30. Fable 5 returned globally on July 1, while Mythos 5 was restored to a set of approved US organizations following government approval on June 26.
OpenAI supplied a second example on June 26 when it previewed GPT-5.6 Sol, Terra, and Luna. At the US government’s request, OpenAI limited the initial release to a small group of trusted partners whose participation had been shared with the government. OpenAI said it did not believe this type of access process should become the long-term default, but proceeded with the restricted preview as a short-term path toward broader availability.
These events do not prove that closed frontier models are unreliable products. They demonstrate something narrower and more consequential for enterprise planning: model access can become conditional.
That turns model dependence from a procurement decision into a business-continuity risk.
Sovereignty Is No Longer Only a National Question
Enterprise AI sovereignty is the set of controls a company maintains over its dependence on intelligence it does not fully own.
Those controls include:
- Where data moves and where inference runs
- Which model handles each class of work
- Whether critical work can move to another qualified model
- Who owns and can inspect the audit trail
- How cost exposure changes at scale
- Whether operations continue during a provider disruption
- Whether the claimed outcome can be independently verified
- How deeply one provider is embedded in the product and workflow architecture
These concerns are not entirely new. Data residency, vendor concentration, disaster recovery, and auditability have long been part of enterprise architecture. What has changed is that the intelligence layer itself can now become restricted, repriced, degraded, or unavailable.
Open Versus Closed Is the Wrong Debate
The argument following these events has largely collapsed into a binary choice between closed frontier models and open-weight alternatives.
That framing is too simple for enterprise deployment, and it is also the wrong decision to be making. Choosing a model category changes where dependency sits, but it does not eliminate dependency.
The enterprise decision is which resolution classes must remain portable across qualified models and which genuinely require a specific model’s capabilities.
The emerging architecture is likely to combine:
- Closed frontier models for high-value reasoning where additional capability justifies the cost and exposure
- Open-weight models for repeatable, private, or cost-sensitive workloads
- Private-cloud or on-premises execution where policy requires a controlled data boundary
- Gateway and routing layers that assign work based on capability, policy, cost, and observability
- Verification infrastructure that determines whether the business problem was resolved
- Independent evidence trails for audit, attribution, and dispute resolution
- Measurement systems that separate AI activity from completed business outcomes
This architecture maps onto CPAG’s Four-Layer Resolution as a Service (RaaS) Stack:
- Physical Edge Runtime
- High-Fidelity Repository
- Agentic Connectivity
- Outcome Interface
Model selection operates inside that stack. It does not replace the stack.
The High-Fidelity Repository is especially important because it provides graph-structured institutional knowledge that agents can use across legacy systems, workflows, and model providers. The repository, not privileged access to one model, is the more durable source of enterprise-specific advantage.
The operating principle follows directly from this: build around the resolution, not the model.
Once a resolution is clearly defined and its full delivery cost is known, controlled model substitution becomes a margin lever. Where another qualified model can meet the same capability, policy, and evidence standards, the vendor can change the delivery input without changing what the customer buys. The resolution and verification standard remain constant while the model underneath can change.
What This Means for SaaS Vendors
Palantir CEO Alex Karp sharpened the commercial problem during a July 1 CNBC interview. He argued that enterprises are paying for tokens that create no value while exposing proprietary data and operating knowledge to model providers. That is Karp’s argument, not proof that every provider behaves this way, but it captures a growing buyer concern: usage is not the same as value.
Karp’s comments also carried a clear commercial interest. Two days earlier, Palantir had announced an expanded NVIDIA partnership centered on deploying open models in sovereign environments. His critique should therefore be read as both a market argument and a positioning argument for Palantir’s alternative.
Vendors that bill on tokens, prompts, messages, or AI credits are exposed in two ways.
First, buyers can question whether metered consumption produced a measurable business result.
Second, controlled model substitution gives buyers an outside option for at least some workloads. A repeatable task may be moved to a less expensive or more controlled model rather than remaining tied to the vendor’s preferred provider.
A better token rate does not answer either concern.
The answer is Resolution Contribution Margin discipline. A vendor must know the direct cost of delivering one verified resolution, including AI compute, data, human oversight, verification, failure costs, onboarding, enablement, and support.
A vendor that understands Resolution Contribution Margin can determine which resolution classes are economically viable, defend pricing, and use model choice to protect margin when a model’s cost or availability changes.
Usage reporting alone cannot demonstrate which resolution classes are economically viable, whether the customer received value, or how model changes affect margin.
This is why consumption pricing is not a bridge to Resolution as a Service. It is the fallback. Consumption billing transfers cost risk to the buyer without proving that a business problem was solved.
Renaming consumption as outcome pricing does not fix the misalignment. The underlying architecture, verification, attribution, and margin model must also change.
What This Means for AI-Native Founders
Founders should treat model portability as a day-one design constraint, even when implementing multiple models on day one would create unnecessary complexity.
A product hardwired to one frontier provider inherits that provider’s pricing, availability, policy, and political exposure. That dependency may be acceptable during initial product development, but it should not be mistaken for a moat.
The more defensible position begins with a High-Fidelity Repository that preserves graph-structured institutional knowledge independently of the model serving it.
Model abstraction, cost observability, resolution definitions, and evidence capture allow a founder to answer a question investors will increasingly ask:
What happens to this business if its primary model becomes unavailable, substantially more expensive, or operationally restricted?
What This Means for Buyers
Before signing or renewing an AI contract, buyers should ask:
- What happens if the primary model is unavailable for 30 days?
- Which resolution classes can move to another qualified model without material degradation?
- Which tasks genuinely require frontier-level capability?
- Can sensitive data remain within the buyer’s approved boundary?
- Can the buyer inspect how a decision or action was produced?
- Is the buyer paying for activity or for a verified resolution?
- Can the vendor show its cost-to-serve and Resolution Contribution Margin by resolution class?
- Who owns the logs and evidence if a regulator, auditor, or customer challenges the result?
- What operational changes are required during a model-substitution event?
These are no longer hypothetical architecture questions. They are procurement and renewal questions.
Model Portability Does Not Eliminate Infrastructure Dependence
Model portability should not be confused with infrastructure independence.
Open weights still require compute, memory, power, integration talent, and operating expertise. The purpose of a model-portable architecture is not to eliminate every external dependency. It is to make those dependencies visible, governed, and replaceable where the resolution’s requirements permit.
The CPAG Position
Enterprise AI strategy is moving from model adoption toward controlled resolution infrastructure.
The durable winners will not necessarily be the vendors with the strongest relationship to today’s leading model provider. They will be the vendors that can:
- Define a valuable Atomic Resolution that is verifiable, attributable, and finite
- Route work across appropriate models
- Preserve enterprise knowledge independently of the model
- Verify that the resolution was completed
- Attribute the outcome to the responsible systems and actors
- Calculate the direct cost of delivery
- Provide the buyer with a defensible evidence trail
The model matters.
The operating architecture determines whether the business can continue delivering the contracted resolution when the model changes.
Prescription
Run a Model Dependency Audit before the next renewal cycle.
Map every critical resolution class currently dependent on one model or provider. For each resolution class, document:
- The primary model
- Qualified alternative models or execution paths
- The required capability threshold
- The acceptable performance degradation
- The data-boundary requirements
- The incremental cost of substitution
- The evidence and audit requirements
- The RCM impact under each delivery option
- Whether the business outcome remains independently verifiable
Any resolution class that cannot survive a 30-day loss of its primary model contains business-model and architectural risk that should be addressed before renewal.
The closing question is simple:
If the vendor could no longer charge for the model call, what verified, attributable, finite resolution would remain valuable enough to charge for?
Sources
- Anthropic, “Claude Fable 5 and Claude Mythos 5”: https://www.anthropic.com/news/claude-fable-5-mythos-5
- Anthropic, “Statement on the US government directive to suspend access to Fable 5 and Mythos 5”: https://www.anthropic.com/news/fable-mythos-access
- Anthropic, “Redeploying Claude Fable 5”: https://www.anthropic.com/news/redeploying-fable-5
- OpenAI, “Previewing GPT-5.6 Sol”: https://openai.com/index/previewing-gpt-5-6-sol/
- Palantir Investor Relations, NVIDIA Nemotron sovereign deployment announcement: https://investors.palantir.com/news-details/2026/Palantir-Launches-Engine-for-Deploying-NVIDIA-Nemotron-Open-Models-in-Sovereign-Environments/
- CNBC, Squawk Box, Alex Karp interview, July 1, 2026