One week, three ways the frontier stopped being freely available
On June 12, days after launch, the US Commerce Department's Bureau of Industry and Security ordered Anthropic to suspend all access to Fable 5 and Mythos 5; the company disabled both models globally within hours. The stated concern was autonomous cyber capability — models that can run multi-step intrusions and discover software vulnerabilities without a human in the loop. On June 26, OpenAI limited GPT-5.6 Sol, its most capable model, to roughly 20 government-vetted partners after a White House request, publicly noting that such restrictions should not become the norm. GPT-5.6 Sol had crossed the "High" threshold on OpenAI's Preparedness Framework and scored 96.7% on its internal capture-the-flag cyber evaluation.
The scaffolding for both moves was the June 2 executive order Promoting Advanced Artificial Intelligence Innovation and Security, which establishes a voluntary pre-release channel — up to 30 days of government access to covered frontier models before public release — and a classified NSA benchmarking process whose thresholds developers cannot see. On June 29, Google's Gemini 3.5 Pro cleared for a July launch as the only major frontier model without an access restriction. A day later, on June 30, Anthropic shipped Claude Sonnet 5 — near-flagship capability at introductory pricing of $2 per million input tokens — as the default model for every free and paid user.
Read together, these are not four disconnected headlines. They describe a single structural shift: the most capable tier of models now reaches government-vetted partners before, or instead of, the market, while the tier just below it flows freely and cheaply. The frontier did not slow down. It split.
Model availability is now a function of national-security review, not vendor roadmap. Architect for the best available model, not the best model.
The Frontier Availability Ladder
The strategic question of the past three years was which model is best. The operating question of this week is which model can you actually run, and how long until you can run the one above it. The Frontier Availability Ladder sorts every model your enterprise might deploy into three rungs by how its availability is governed. The single metric that spans the ladder is capability latency: the elapsed time between a capability existing at the frontier and your organisation being cleared to deploy it in production.
The most capable models, released first or only to government-vetted partners. Availability is governed by national-security review, export-control action and classified thresholds. GPT-5.6 Sol, Claude Fable 5 and Mythos 5 live here today. For any enterprise outside the trusted-partner circle — which is nearly every enterprise outside the US perimeter — this rung is a capability you can read about but not run.
Near-flagship models that are publicly available after — or without — review. Gemini 3.5 Pro cleared with no restriction; Claude Sonnet 5 lands close to flagship Opus 4.8 on reasoning, coding and tool use, at a third of the cost. Availability is governed by vendor roadmap plus clearance timing. This is the rung on which the productive economy actually runs, and where most enterprise workloads belong by design.
Models you host yourself or draw from regional public infrastructure — open weights, and sovereign efforts such as Chile's Latam-GPT at CENIA. Availability is governed by your own infrastructure and cannot be revoked by another government. Capability lags the frontier, but the latency is bounded and knowable. This rung is the only one whose clock you control.
So what: the procurement question is no longer "which model is best". It is "which rung serves this decision, what is my capability-latency tolerance for it, and what do I fall back to when a rung above closes without warning?"
Three LATAM operators, three availability postures
CABA Tier-1 bank — fraud and AML on a broken assumption. A Buenos Aires universal bank had architected an autonomous fraud-analytics workload on the assumption of continuous access to a flagship model; the June 12 global suspension of Fable 5 broke that continuity overnight, mid-quarter. The rebuild routes the workload to a Rung 2 cleared-frontier primary (Claude Sonnet 5) with a Rung 3 sovereign fallback (Latam-GPT/CENIA) for regulated inference under Ley 25.326. Cost per decision falls 38%; substitution latency holds under 30 days; sovereign-substrate coverage on regulated flows reaches 34%; human-in-the-loop override stays under 8%.
São Paulo industrial logistics — forecasting on a locked model. A Brazilian logistics operator had piloted dock-scheduling and demand forecasting on a GPT-5.6-class capability that is now confined to government partners. Rather than wait on an unknowable clearance clock, it falls back to a Rung 2 primary (Gemini 3.5 Pro) with a Rung 3 Portuguese-first sovereign substrate under LGPD. Mean time to recovery on the scheduling agent drops from 41 min to 9 min; on-time-in-full improves 8.4pp; a capability-latency tolerance of ≤45 days is written into the workload SLA so no single locked model can stall operations.
Multi-country LATAM grain exporter — designing for permanent Rung 2. A fourteen-port exporter with operations across Argentina, Paraguay and Uruguay assumes it will never sit inside the trusted-partner perimeter and architects accordingly: a Rung 2 cleared-frontier tier for commercial workloads and a Rung 3 sovereign tier for customs, sanctions-screening and citizen-facing flows. It treats Rung 1 as strategically inaccessible rather than delayed. Sovereign-substrate coverage on regulated workloads reaches 42%; identity-attested action holds at 100%; portfolio inference cost falls 41%; model concentration stays under 60% per provider.
Implementation: a capability-latency budget, not a wish list
The enterprise that survives this shift does not maintain a list of the models it wants. It maintains a capability-latency budget: for each production decision, the rung that serves it, the maximum tolerable delay before an upgrade, and the fallback that keeps the lights on when a rung above closes without notice.
Rung 1 gave the market a lesson in June: the newest model is a privilege, not a purchase. The rows of your budget are still your decisions; the columns are now the three rungs and the clock between them.
So what: from pilot to policy. KPIs before APIs. Interoperability or it doesn't scale. Availability is the new line item on the operating-layer scorecard.
Governance
Add a model-availability continuity clause to every frontier contract: notice periods, substitution rights, and portability of prompts, evals and fine-tunes. Map regulated decisions to Rung 3 by default. Track the US executive order's voluntary review channel, BIS export actions, EU AI Act systemic-risk provisions, Ley 25.326 and LGPD as one moving perimeter. Treat a mid-contract gating event as a modelled risk, with a human-in-the-loop failover owner named in advance.
KPIs
Capability-latency tolerance defined per workload, in days. Cleared-frontier coverage of production traffic. Sovereign-substrate coverage on regulated workloads above 40%. Substitution latency under 30 days. Model concentration below 60% per provider. Cost-per-decision delta of at least 35% versus baseline. Override rate under 8%. Decision auditability at 100%.
12-month roadmap
0–90: map every production decision to its rung, tag single-model dependencies, set a capability-latency tolerance per workload. 90–180: stand up a model gateway with a Rung 2 primary and Rung 3 sovereign fallback, then run a 30-day gating-shock failover drill. 180–360: reach ≥40% sovereign coverage on regulated flows and report a capability-latency scorecard to the board each quarter.
The best model is the one you are allowed to run.
For three years the strategic conversation assumed the newest, most capable model would be available to anyone who could pay for it. This week ended that assumption. When a government can order a model disabled worldwide within hours, and when the most capable release ships first to twenty vetted partners, availability itself becomes a design constraint — and for enterprises outside the US perimeter, a structural one. Latin American operators do not sit inside the trusted-partner circle, and it is prudent to assume they never will. That is not a disadvantage to lament; it is a spec to build against.
The operators who will absorb this without disruption are the ones already treating the cleared frontier as their production floor and a sovereign substrate as their insurance — Latam-GPT and comparable regional infrastructure not as a symbolic gesture but as the one rung whose clock they control. Claude Sonnet 5 shipping at a third of flagship cost is the quiet good news underneath the gating headlines: the capability the productive economy actually needs is cheap, abundant and unrestricted. The scarce thing is not intelligence. It is the discipline to architect for the model you can run today and the latency budget to reach the next one. From pilot to policy. Build for the cleared frontier, and own the rung beneath it.
Build your capability-latency budget
Socradata runs Frontier Availability Diagnostics for LATAM operators in financial services, logistics, agribusiness and the public sector. The output is a decision-by-rung map, a capability-latency budget, a sovereign-substrate fallback plan and a governance overlay with model-availability continuity clauses ready for board review.