I read the Fable 5 shutdown as a contract failure before I read it as a policy event. The directive was outside any customer's control. The contract was not.
On June 12, 2026 at 5:21pm ET, a US Commerce Department directive ended Fable 5 and Mythos 5 access for enterprise customers globally in 90 minutes. Core AI services went dark mid-workflow. The Department of Defense, which had moved significant agentic workloads to Fable 5 after the June 9 launch, lost them. Finance and healthcare teams that had done exactly what a prudent organization should do — evaluated the model, ran evals, migrated production workloads — found themselves in the same position as teams that had done nothing: capabilities gone, no notice, no contractual remedy.
I have advised boards and executive teams on technology risk for twenty years. I have seen cloud outages, vendor acquisitions, and regulatory changes take down services before. What happened on June 12 was different in one specific way: the gap was not in the infrastructure; it was in the contracts. Standard SLAs, force majeure language, and notification duties were not designed for a government-directed model capability restriction, and they failed at every point where they were needed.
The three contract assumptions that failed are specific, and each one is fixable. Naming them precisely is where the corrective starts.
The three assumptions the contract made — and the directive falsified
The first assumption: that unavailability means the endpoint stopped responding. Every standard SLA defines uptime as infrastructure availability — the API answers, the latency is in bounds, the region is healthy. On June 12, all of that was fine. The infrastructure never failed. The capability was simply withdrawn by directive, and the SLA had no language for a model that exists and is not served. Enterprise customers held technically-compliant agreements with a vendor whose infrastructure was fully operational, for a model they could not use.
The second assumption: that force majeure is a shared shield. It is not, and reading one carefully makes this obvious. Standard force majeure clauses are drafted by vendors to excuse the vendor's own non-performance, and they routinely enumerate government action as an explicitly excluded event — meaning the service-credit obligation evaporates the moment a government acts. The clause that reads, to a procurement team, as mutual protection against acts of God is in practice a clause that released Anthropic from any credit obligation while leaving the customer holding a disabled workflow. The asymmetry is by design. June 12 made it expensive.
The third assumption: that the vendor owes the customer warning. Most AI contracts impose no notification duty on the vendor for regulatory events affecting the service. Anthropic received the Commerce directive at 5:21pm ET. Enterprise customers learned of the cutoff not from their vendor but from their own incident channels lighting up as production calls failed. There was no notification duty, no cure period, no clock the customer could plan against. A capability priced at $10 per million input tokens and $50 per million output, and built into production workflows at that price, was gone before most incident teams had read the first alert.
Six clauses that close the gaps
These are drafting instructions, not principles. Each clause names the specific gap the Fable 5 event exposed and the specific language that closes it. I would bring them to every enterprise AI contract renewal as line items, not as a wish list. The precedent is documented and the vendor now has an incentive to differentiate on exactly these terms.
Capability availability definition
The contract must define availability at two distinct layers and assign an obligation to each. Infrastructure uptime — the endpoint responds — is the layer every current SLA already covers. Capability availability — the specific contracted model produces output to the contracted spec — is the layer that failed on June 12 and that no standard SLA names. Require the agreement to enumerate each model in scope by name and version and to attach the service-credit schedule to capability availability, not only to endpoint response.
Government-directed restriction notification
The vendor must notify the customer within a defined window — 24 hours is the workable floor — of any government directive that restricts, suspends, or withdraws contracted capabilities, to the extent disclosure is itself lawful. On June 12 customers learned of the cutoff from their own incident channels lighting up, not from Anthropic. A notification clause does not prevent the directive; it converts a silent mid-workflow failure into a managed incident with a clock the customer can plan against.
Tested-fallback representation
The vendor warrants that named alternative models have been validated on the customer's own workloads and are available at no additional cost when primary capabilities are restricted. This is a representation, not a best-efforts promise: the fallback model is named in the contract, the validation evidence is attached, and the switch path is pre-agreed. A fallback that has never run your evaluation suite is not a fallback — it is a hope, and June 12 is what hope costs.
Customer-side force majeure
Standard force majeure protects the vendor: government action excuses the vendor's performance and ends the service-credit obligation. The corrective is a mirror clause that runs the other way. If a government-directed action makes contracted capabilities unavailable for more than a defined period (7 days is a defensible threshold for critical workloads), the customer may terminate without penalty and receive a pro-rata credit for the unused term. The same event that excuses the vendor must also release the customer.
Data portability guarantee
Embeddings, fine-tuned weights, prompt libraries, and evaluation artifacts must be exportable in a usable format within 30 days of termination, with the format and delivery mechanism specified in the contract rather than left to the vendor's discretion at exit. A capability shutdown is survivable if you can move the workload; it is catastrophic if the assets that make it portable are stranded behind a provider that can no longer serve them.
Regulatory risk disclosure
The vendor must notify the customer of any government investigation, export-control proceeding, or licensing action that may affect the service, within 48 hours of vendor receipt. Most current AI contracts impose no duty to disclose regulatory exposure to the service the customer depends on. After Fortune framed the shutdown risk as a standing disclosure category for AI vendors, the absence of this clause is no longer an oversight — it is an accepted information asymmetry the customer is choosing to keep.
The board disclosure question
The Fable 5 case created a risk category that most boards had not named. AI vendor risk had been filed, where it was filed at all, under technology risk generally — alongside cloud outages and dependency patching. June 12 separated it out. The new line item is specific: "US export control authority over AI provider." It is not a hypothetical. It is a demonstrated mechanism by which a single government action can remove a contracted capability globally, instantaneously, with no notice and no remedy.
A risk register that does not carry this as a distinct entry is now incomplete. And the reporting obligation follows from the risk. Any material dependence on a US-controlled frontier model for critical operations belongs in board reporting with two attachments: a documented contingency plan and an independence test result — the actual measured ability to move the workload to a named alternative, not the assertion that one exists. A board that receives a "we use AI for everything" summary without a sovereignty and resilience annex is receiving a risk picture with the most expensive variable left out.
Using the Fable 5 case as negotiating leverage
Procurement teams negotiating now hold a stronger hand than they did on June 11. The risk is no longer a thought experiment a vendor can wave away; it is a documented event with a date, a directive, and a Fortune headline. That changes the conversation from "could this happen" to "here is what happens, and here is the contractual protection that distinguishes you from the provider it happened to." A vendor that wants the deal now has an incentive to offer exactly the clauses Anthropic's contracts lacked.
The concrete ask: for current renewals, shorten the initial term — twelve months with an option to extend, not a multi-year lock — and attach explicit capability availability representations for each model in scope. For new relationships, lead with the tested-fallback warranty and the customer-side force majeure clause. A vendor confident in its regulatory posture should have no objection to either. A vendor that resists them is telling you something. Above all, do not sign a long-term enterprise agreement that assumes Fable 5 returns on a predictable timeline. Anthropic and Commerce are in a multi-month negotiation with no committed date and no public timeline. Pricing a multi-year commitment against a capability currently subject to export controls is the mistake the June 12 cohort is now unwinding.
The position
Every enterprise that found a core capability disabled mid-workflow on June 12 had signed an agreement that defined availability as endpoint uptime, treated force majeure as the vendor's shield, and asked nothing of the vendor by way of warning. None of that was inevitable, and all of it is fixable in the next negotiation.
My recommendation is narrow and I would put it in writing to any board I advised: treat June 12 as the moment AI vendor risk became a named, disclosable, contractually addressable category, and act on it before the next renewal rather than after the next directive. A capability shutdown is survivable when the contract anticipated it. It is a crisis only when the contract assumed it could never happen.
What happened to enterprise AI contracts when Fable 5 was shut down?
Enterprise clients with active Anthropic agreements — including Fable 5 and Mythos 5 in their contracted capability set — found those capabilities unavailable within 90 minutes of the US Commerce directive on June 12, 2026. Standard SLAs define uptime obligations for infrastructure availability, not model capability availability. Government-directed model withdrawals fall outside the typical uptime calculation entirely and are usually carved out of force majeure provisions. The result: enterprise customers with active service agreements had no contractual remedy for the loss of capabilities they had contracted and planned around.
Does force majeure cover a government-directed AI model shutdown?
In most current enterprise AI contracts, yes — and that is the problem. Standard force majeure clauses are designed to protect vendors, not customers, and typically enumerate government actions as explicitly excluded from SLA obligations. The clause that was supposed to protect both parties from acts of God became the clause that protected Anthropic from any service credit obligation to enterprise customers. After June 12, the corrective is to add a customer-side force majeure provision: if a government-directed action makes contracted AI capabilities unavailable for more than X days, the customer has a right to terminate without penalty and receive a pro-rata credit for unused contract value.
What should AI vendor contracts include after the Fable 5 case?
Six additions. One: a capability availability definition that distinguishes infrastructure uptime from model capability availability. Two: a government-directed restriction notification clause. Three: a tested-fallback representation — the vendor warrants that named alternative models have been validated on the same workloads. Four: customer-side force majeure. Five: data portability guarantee — embeddings, fine-tuned weights, and evaluation artifacts exportable within 30 days. Six: regulatory risk disclosure — the vendor notifies customers of any government proceeding affecting the service within 48 hours of receipt.
Can enterprises recover damages from Anthropic for the Fable 5 shutdown?
Unlikely under most current contracts. The combination of standard force majeure carve-outs for government actions, SLA definitions limited to infrastructure uptime, and limitation-of-liability caps (typically capped at fees paid in the preceding 3–12 months for direct damages, excluding consequential damages entirely) means enterprise customers have limited legal exposure to recover. The practical remedy was operational, not legal: identify the workloads that failed, redirect to available alternatives, and document the business impact for future contract negotiation leverage.
Should we sign new long-term enterprise AI contracts right now?
Not until the contract includes the six clauses above, and not on a model currently subject to export controls. Do not sign multi-year contracts assuming Fable 5 returns on a predictable timeline — Anthropic and Commerce are in a multi-month negotiation with no public timeline. For current renewals, negotiate shorter initial terms (12-month initial, option to extend) with explicit capability availability representations for each model in scope. For new vendor relationships, use June 12 as leverage: the risk is documented, the precedent is set, and the vendor has an incentive to offer protections that distinguish them.
What is the board's responsibility after the Fable 5 case?
Two things. First, AI vendor risk should appear on the board's risk register as a distinct line item — not buried in technology risk generally, but named: "US export control authority over AI provider." The Fable 5 case created a new risk category most boards had not mapped. Second, any material dependence on a US-controlled frontier model for critical operations should be disclosed in board reporting with a documented contingency plan and an independence test result. Boards that receive a "we use AI for everything" report without a sovereignty and resilience annex are receiving an incomplete risk picture.
Frequently Asked Questions
What happened to enterprise AI contracts when Fable 5 was shut down?
Enterprise clients with active Anthropic agreements — including Fable 5 and Mythos 5 in their contracted capability set — found those capabilities unavailable within 90 minutes of the US Commerce directive on June 12, 2026. Standard SLAs define uptime obligations for infrastructure availability, not model capability availability. Government-directed model withdrawals fall outside the typical uptime calculation entirely and are usually carved out of force majeure provisions. The result: enterprise customers with active service agreements had no contractual remedy for the loss of capabilities they had contracted and planned around.
Does force majeure cover a government-directed AI model shutdown?
In most current enterprise AI contracts, yes — and that is the problem. Standard force majeure clauses are designed to protect vendors, not customers, and typically enumerate government actions as explicitly excluded from SLA obligations. The clause that was supposed to protect both parties from acts of God became the clause that protected Anthropic from any service credit obligation to enterprise customers. After June 12, the corrective is to add a customer-side force majeure provision: if a government-directed action makes contracted AI capabilities unavailable for more than X days, the customer has a right to terminate without penalty and receive a pro-rata credit for unused contract value.
What should AI vendor contracts include after the Fable 5 case?
Six additions. One: a capability availability definition that distinguishes infrastructure uptime from model capability availability. Two: a government-directed restriction notification clause. Three: a tested-fallback representation — the vendor warrants that named alternative models have been validated on the same workloads. Four: customer-side force majeure. Five: data portability guarantee — embeddings, fine-tuned weights, and evaluation artifacts exportable within 30 days. Six: regulatory risk disclosure — the vendor notifies customers of any government proceeding affecting the service within 48 hours of receipt.
Can enterprises recover damages from Anthropic for the Fable 5 shutdown?
Unlikely under most current contracts. The combination of standard force majeure carve-outs for government actions, SLA definitions limited to infrastructure uptime, and limitation-of-liability caps (typically capped at fees paid in the preceding 3–12 months for direct damages, excluding consequential damages entirely) means enterprise customers have limited legal exposure to recover. The practical remedy was operational, not legal: identify the workloads that failed, redirect to available alternatives, and document the business impact for future contract negotiation leverage.
Should we sign new long-term enterprise AI contracts right now?
Not until the contract includes the six clauses above, and not on a model currently subject to export controls. Do not sign multi-year contracts assuming Fable 5 returns on a predictable timeline — Anthropic and Commerce are in a multi-month negotiation with no public timeline. For current renewals, negotiate shorter initial terms (12-month initial, option to extend) with explicit capability availability representations for each model in scope. For new vendor relationships, use June 12 as leverage: the risk is documented, the precedent is set, and the vendor has an incentive to offer protections that distinguish them.
What is the board's responsibility after the Fable 5 case?
Two things. First, AI vendor risk should appear on the board's risk register as a distinct line item — not buried in technology risk generally, but named: "US export control authority over AI provider." The Fable 5 case created a new risk category most boards had not mapped. Second, any material dependence on a US-controlled frontier model for critical operations should be disclosed in board reporting with a documented contingency plan and an independence test result. Boards that receive a "we use AI for everything" report without a sovereignty and resilience annex are receiving an incomplete risk picture.
Ready to Transform Your AI Strategy?
Get personalized guidance from someone who's led AI initiatives at Adidas, Sweetgreen, and 50+ Fortune 500 projects.
