Frontier model launches now run through Washington

Anthropic's 18-day shutdown and OpenAI's vetted preview show that release timing at the frontier is now a policy variable, not an engineering one.
Anthropic's newest models, Fable 5 and Mythos 5, spent 18 days offline before the Commerce Department withdrew its export controls on 30 June. The order had arrived on 12 June, three days after Fable 5's public launch, and barred access by any foreign national. Unable to check nationality in real time, Anthropic switched both models off worldwide.
The launch date of a frontier model is now set partly in Washington, and capital has to price that in. Four days before the controls lifted, OpenAI opened GPT-5.6 only to a small group of trusted partners approved by the US government, at the government's request. Within a month, government action shaped launches at both labs, one through export controls and one through a review process. Revenue from a new model now starts later and less predictably, and a single safety finding can take a flagship product off the market overnight.
That review process comes from the executive order of 2 June, which created a voluntary path for the government to test covered frontier models for up to 30 days before wider release and to help pick early-access partners, a step beyond the 2023 reporting order. Anthropic's way back ran through commitments of the same kind. Access returned from 1 July after it agreed to detect and address security risks proactively and to coordinate with the government on protocols for future releases. Absorbing such a process takes security teams and standing ties to government, which favours a few incumbents over new entrants.
A product built on day-one access to one frontier model carries that risk in full, a lesson OpenAI's 2023 crisis already taught. Anthropic's other models stayed online through the shutdown, and customers able to fall back to them kept working. Pricing makes the same case from the other side. GPT-5.6 Sol lists at $5 per million input tokens and $30 per million output tokens, while the smaller Luna costs $1 and $6, so a product that sends routine steps to the cheaper tier keeps its margin even when the top tier is gated. What the government asked for also opens a market in jailbreak detection, misuse monitoring and audit tooling that proves safeguards work.
One month of interventions could still prove an exception rather than a regime. The order asks the government to set a threshold for covered frontier models, and that line decides which labs fall inside the process. GPT-5.6 is the nearer test: if the gap between limited preview and general availability runs to weeks, gated launches become the norm. The longer test is whether the commitments Anthropic made to regain access become a template that other labs sign before they ship.
Fanaura invests as long-horizon capital, and over that horizon the shift is structural. Access to the best models now carries regulatory risk alongside price and performance, and that risk lands hardest on companies that depend on one supplier. We favour builders whose value sits in their data, workflow and distribution. In their products any single model remains a replaceable input, and a pause in Washington becomes a routing decision rather than an outage.
Building an AI-native company? Apply in five minutes. →
More from Fanaura
Building tools that let AI products switch models or prove their safeguards? Send us the project.
Fanaura invests its own capital and the capital of its limited partners (LPs), usually through a separate SPV for each deal. If we see a fit, we get in touch.
Get the news by email
One short note each weekday. No marketing, and you can unsubscribe at any time.