When you buy an AI model, the routine is familiar: check the certifications, sign the SLA, pick the best benchmark and move on. The quiet assumption is that availability is a technical problem with a contractual answer.
In June 2026 that assumption broke. Anthropic disabled Claude Fable 5 and Mythos 5 worldwide on roughly 90 minutes’ notice, and OpenAI withheld GPT-5.6 behind a government gate. Compliance protected nobody; the switch was thrown outside any buyer’s contract.
Before you compare benchmarks, ask who can switch it off. Here’s the full picture behind AI vendor regulatory risk.
How should you assess an AI vendor’s regulatory exposure, and whether a provider can lose access to its own product overnight?
Score a vendor on jurisdiction of control, export-control exposure, gatekeeper dependence and contractual durability, alongside certifications. Could a directive or a Trusted Partners gate revoke access with little notice? You want a scored exposure profile you can feed into procurement.
Why is AI vendor regulatory risk really an availability risk?
Traditional vendor-risk planning treats failures as technical and temporary. Model Access Revocation is neither: access is withdrawn with no technical fault and no uptime breach. June 2026 was the first large-scale case, when Anthropic said it had to “abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance”.
What should a regulatory-exposure scorecard measure?
A regulatory-exposure assessment scores exposure on a graded scale rather than as a pass/fail stamp. Each instrument fails differently, so score them separately: a gated launch withholds a model at release, an export-control designation recalls it after launch, and a procurement condition overrides commercial terms. Add concentration, three providers hold roughly 88% of production AI usage, and check for an open-weight fallback. Trusted Partners, the mechanism that decides who keeps access when a model is gated, matters as much as any clause. See how OpenAI and Anthropic responded. Scoring exposure is only half the job; the gate also travels through the vendor’s contract.
How should you assess the procurement and compliance risk of a government-gated model?
Check whether the model is gated at launch or after deployment, and who qualifies as a Trusted Partner. Then trace the vendor’s federal obligations downstream, because a GSA-style clause can override commercial terms. Require regulatory-flexibility, exit, data-return and model-change notice up front.
How does a US-federal obligation reach a customer who never sells to the US government?
Under the government gatekeeper model, access is pre-cleared by government, general availability is withheld, and trusted partners are admitted first. That gate travels by procurement-as-governance: the vendor’s federal obligations become your obligations through the supply chain. The GSA’s new AI clause defines a Service Provider as any entity that indirectly provides AI to a contractor, even with no government contract. The clause overrides commercial terms, so non-compliant invoices can become False Claims Act claims. Deemed export adds another vector: a vendor’s own foreign-national employees can trigger licensing, so a staffing decision upstream affects your access.
What contract terms and exit clauses should your business require?
For a material AI vendor, negotiate model-change notice with version pinning, data-return and exit clauses, audit rights and indemnity. Standard indemnities rarely cover model withdrawal, so spell that out. Define customer data to include fine-tuned weights, embeddings, agent memory and conversation traces, because those exclusions make the system portable. Then add event-driven reassessment: a material model change, an incident or a lost certification should trigger review. That is the pre-release gate that creates the overnight-loss risk. If access can end overnight, the next question is what you hold when it does.
Government-gated closed models vs open-weight models: which is the safer long-term bet for your business?
A government-gated closed model offers frontier capability, but access is revocable by a decision outside your control. An open-weight model can be downloaded and self-hosted, so it cannot be remotely switched off. The trade-off is that the largest open models exceed most buyers’ serving capacity. The safer bet depends on process criticality and jurisdiction, which capability alone cannot settle.
What does an open-weight model actually protect you from?
Closed models keep their weights locked away, reachable only through a provider’s interface. Open weights let you download and run the model yourself, turning AI from a rented service into shared infrastructure with no upstream left to fail; once released, weights cannot be recalled. Open weights trail the frontier by roughly six months and are not open source either; training data and methods stay private.
How much does Australia’s distance from Washington change the bet?
The short answer for Australian buyers: less than they might hope. US export-control reach follows the vendor and its obligations rather than the customer’s postcode. Data residency does not stop a jurisdictional access cut; the US Cloud Act can compel American providers to hand over data held abroad. Australia has no credible path to self-sufficiency, so the realistic position is managed dependence rather than autarky. The Technology Prosperity Deal and lab agreements seek durable access rather than independence. That is how rival governance regimes change the hedge. The availability question also changes how you read the capability race, starting with Sol versus Mythos.
GPT-5.6 Sol vs Anthropic Mythos 5: which is more capable, and on what tasks?
No independent apples-to-apples benchmark exists; every figure traces to vendor reporting. Scope by task. GPT-5.6 Sol leads on broad coding and agentic work, while Mythos 5 leads on cyber and exploitation-style tasks. Treat the Mythos/Fable cyber score as conflated. Choose on task fit and continuity risk rather than a single leaderboard number.
GPT-5.6 Sol vs Terra vs Luna: which tier should you actually choose?
GPT-5.6 ships as Sol (flagship), Terra (balanced) and Luna (fast, cheap), surfaced through ChatGPT, Codex and the API. Pick the smallest tier that clears your accuracy bar and reserve Sol for the hardest agentic jobs, since the Sol-to-Luna price gap is 25x while the benchmark gap is narrower, making Terra the default at less than half the price.
Sol sets state of the art on BrowseComp and OSWorld 2.0, and on Terminal-Bench 2.1 Sol Ultra reaches 91.9% against Fable 5’s 86.0%. On SWE-Bench Pro, Fable 5 leads at 80% to Sol’s 64.6%, and OpenAI claims 30% of those tasks are broken. The 78% ExploitBench figure belongs to Mythos 5 rather than Fable 5. Fable 5 and Mythos 5 share an architecture and differ only in safeguards, so the name signals the access tier rather than the underlying capability.
The frontier is split; anyone saying one lab swept the board is selling something. Verify before citing, then weight the availability question into the verdict, and keep a fallback you hold for the day neither is reachable. The OpenAI and Anthropic cases condition that verdict.
Vendor choice is a bet on who holds the switch. Score exposure on the four dimensions above, keep an open-weight fallback and a multi-model strategy you control, and read the verdict as task-scoped and continuity-weighted: Sol for coding, Mythos for cyber. That is the broader shift.
Frequently Asked Questions
Can my AI provider really switch off my model overnight?
Yes, if the model is government-gated or subject to export controls. In June 2026, Anthropic disabled Claude Fable 5 and Mythos 5 worldwide on roughly 90 minutes’ notice, and OpenAI withheld GPT-5.6 behind a government gate. The switch was thrown by a decision outside any customer contract, so an uptime SLA offered no protection at all.
Is open-weight the same as open source?
No. Open-weight models release the trained parameters so you can download and self-host them, but the training data, code, and methods usually stay private. That still matters for continuity: because the weights sit on your infrastructure, no vendor or government can remotely revoke them. It does not give you the transparency or the right to retrain that true open source implies.
What is a Trusted Partner in AI procurement?
A Trusted Partner is an organisation pre-cleared by government to keep access to a gated model when general availability is withheld. Under the government gatekeeper model, access is granted selectively to these partners first, so your place on that list, not your contract, may decide whether you keep the model. Confirm your tier before you build on a gated release.
Does the US Cloud Act apply to data stored in Australia?
Yes. The US Cloud Act can compel US-based providers to hand over data they hold, even when it is stored on Australian servers. That means data residency alone does not prevent a jurisdictional access cut. If the provider is subject to US law, the location of the servers matters less than who ultimately controls the service.
What is deemed export, and how does it affect AI vendors?
A deemed export occurs when foreign-national employees access controlled technology inside the United States, which can trigger a licensing requirement even though nothing is physically shipped. For AI vendors this means their own workforce can create export-control exposure. It also means your risk is not limited to your own country, because a vendor’s staffing decision upstream can affect your access.
Is AI sovereignty the same as data residency?
No. Data residency is about where data sits; AI sovereignty is about whose legal system governs the model and who can switch it off. You can run a model on your own infrastructure for residency and still depend on a foreign vendor for weights, updates, and licences. The realistic position is managed dependence rather than full self-sufficiency.
What is Model Access Revocation?
Model Access Revocation is access being withdrawn independently of any technical fault or uptime promise. It can arrive three ways: a gated launch where a model is withheld at release, an export-control designation that recalls it after launch, or a procurement condition that overrides commercial terms. Each fails differently, so each should be scored separately in a vendor assessment.
What happens to my data and models if my AI vendor’s access is revoked?
That depends entirely on your contract. Without explicit exit and data-return terms, you may lose access to the model, your fine-tunes, and your logs with little notice. Standard indemnities rarely cover model withdrawal, so negotiate data return, model-change notice, and version pinning before you build. Plan on the assumption that access can end abruptly.
Is an uptime SLA enough to guarantee model availability?
No. An uptime SLA only promises the service stays technically available; it says nothing about a government directive or a Trusted Partners gate that revokes access outright. June 2026 showed availability can be withdrawn outside any contract, so score contractual durability and jurisdiction of control alongside the SLA rather than treating it as protection.
GPT-5.6 Sol vs Terra vs Luna: which tier should a team really choose?
Choose by task and token cost, not raw capability. Match the smallest tier that clears your accuracy bar for the work in front of it, then reserve the top tier for the hardest reasoning and agentic jobs. Because agentic tasks consume far more tokens, the tier you pick drives cost-per-unit-of-work more than any headline benchmark does.
What are the early warning signs that a vendor is about to be gated?
Watch for a model held back at launch, new export-control or supply-chain designations, and tightened Trusted Partner tiers. A vendor that cannot confirm its foreign-national staffing exposure, or that shifts to selective access, is signalling gate risk. Treat any of these as a trigger to reassess and to confirm your open-weight fallback is ready.
Does being in Australia protect me from US AI export controls?
Only partially. Distance from Washington does not insulate local buyers, because US export-control reach follows the vendor and its obligations, not your postcode. Australia’s Technology Prosperity Deal and lab agreements seek durable access rather than true independence, which still leaves the switch upstream. Keep a fallback you control even if your access looks secure.
