Vendors · Regulation
Frontier launch dates,
now on Washington's calendar.
Today's GPT-5.6 release and this week's Anthropic export controls look like separate stories. Read together, they map to the same shift: for frontier-class models, the ship date is set by US government review, not by the lab. Enterprise roadmaps just got one more line item to carry.
Two Data Points
Two dots on the same map
Today's GPT-5.6 clearance isn't a standalone story. Under the per-customer approval rule introduced in late June, frontier-class models now ship one at a time, only after the US government signs off. GPT-5.6 is the first dot on that map, and this week's Anthropic model export controls is the second.
Line up the two, and the picture is unmistakable. Frontier-class launch schedules are now effectively set by the US government. That's not a per-lab or per-quarter observation — it's a shift in the industry's cadence itself.
The Two Gates
Two gates that hold a launch
Approvals come in two stages. A model can stall at either.
Model-level gate
The lab submits the model's benchmarks and intended use, and asks whether this generation can ship at all. Fail here and the release simply doesn't move — GPT-5.6 sat here for three weeks.
Customer-level gate
Even after the model clears, each buyer's country and company is re-reviewed. "Sellable / not sellable" flips by jurisdiction — some Anthropic models were stopped at this stage for third-country customers this week.
What's New
The new roadmap line
What To Watch
What to keep on your dashboard
The load falls almost entirely on the person owning model selection inside a company. For a person just chatting with an AI, essentially nothing changes. Enterprise procurement should keep three lines in mind.
1) Assume next-gen models (GPT-5.7, the Claude 4.1 tier, Gemini 2.5 Pro tier) will hit the same two gates — loosen cutover SLAs by roughly one government-review window. 2) List a fallback model at contract time for every country where "yes / no" could split. 3) Push Go/No-Go dates on quarterly roadmaps a bit further back to absorb slippage.
Today's GPT-5.6 restart provides the first real data point for this new rhythm. Three weeks — long enough to matter, short enough to plan around — is a reasonable margin to start assuming.