Who decides which AI your agency can use

A new OpenAI model shipped this week. GPT-5.6 Sol beats Claude Opus on several key benchmarks and sits at the top of the frontier tier by most credible measures. For most agencies, that would normally be straightforward news: a better tool is available, go evaluate it, update your stack.

Except this time, the US government is deciding who gets access.

Sol launched with tiered availability tied to export controls and bilateral agreements between the United States and partner countries. Certain regions, certain corporate structures, certain use cases: all gated. Not by OpenAI's pricing page, but by federal policy. That is a different kind of constraint than anything the creative and product industry has dealt with before from a software vendor.

This has been coming for two years

The US government has been treating frontier AI models as strategic assets since at least 2023. The Commerce Department's Bureau of Industry and Security extended export control logic from chips to model weights and API access. The executive orders from both the Biden and Trump administrations, despite their differences in tone, agreed on one thing: the most capable AI systems should not flow freely to adversarial states.

What is new with Sol is that the controls are visible at the product level in a way they weren't before. Earlier access programs were quiet agreements, early-access waitlists, or geography-based API blocks that felt like operational logistics. Sol makes the government's hand explicit. OpenAI has said publicly that access for certain regions requires US government clearance as part of the deal structure.

This is not unprecedented in tech. Encryption software was export-controlled for decades. Satellite imagery has always had government strings attached. What's different here is speed. The policy apparatus is moving faster than most organisations have updated their AI procurement thinking.

What this means if you run an agency

Let's be specific about the operational pressure here, because the abstract geopolitical framing can make it feel distant.

First. your model stack now has a country-of-origin problem. If you're a European agency working for a client with operations in a restricted jurisdiction, the model your engineers rely on may not be available for that project. That's a workflow gap you need to plan for, not react to.

Second. the gap between frontier and second-tier models matters more than it used to. Sol outperforms its nearest open-weight alternatives on reasoning and code tasks by a meaningful margin. If access to it is gated, you're not just choosing a different vendor, you're accepting a capability downgrade on specific task types. That affects what you can promise clients and how long certain work takes.

Third. open-weight models are back on the table as a serious option. Llama 3.3, Mistral's Magistral, and others are genuinely capable for a wide range of agency work. They're not restricted in the same way. If you dismissed them six months ago because the closed frontier models felt more capable and easier to operate, the access-control picture gives you a reason to revisit that calculation.

Fourth. your clients will ask you about this. Not immediately, but soon. Procurement teams at larger companies are already fielding questions from legal and compliance about which AI vendors are permissible under their data agreements. When the client's question reaches you, you want an answer ready.

The contrarian case for staying calm

There is an argument that this is smaller than it looks, and it is worth taking seriously.

Most agency work, most of the time, does not require the single best available model. It requires a model that is good enough for the task, integrated reliably into a workflow, with costs that don't blow the project margin. GPT-4o, Claude 3.7 Sonnet, and Gemini 1.5 Pro are all still fully accessible. They are all genuinely capable. For the majority of content, design research, code assistance, and data tasks that agencies actually do, the frontier gap is academic.

The risk of overcorrecting here is real. Agencies that spend the next quarter auditing geopolitical AI risk instead of building good work for clients will fall behind on both counts. The models that are available to you right now are better than what you had twelve months ago by a significant margin. That should stay the center of gravity.

Still, ignoring the structural shift entirely is a mistake. The question is not whether to panic. It is whether to think about model access as a static product decision or as something that now has a policy dimension you need to track.

What a studio like ours does with this

We've run multi-model architectures for over a year now. Not because we are hedging politically, but because no single model wins every task category. Routing creative research through one model, structured data extraction through another, and code generation through a third has been standard practice for us since mid-2024. That setup, built for capability reasons, turns out to be resilient to exactly this kind of access disruption.

The implication for any serious agency. model agnosticism is now a business continuity position, not just a performance optimization. Build your workflows so that the model layer is swappable. Keep your abstraction one level up from any specific vendor's API. If you are deeply coupled to a single provider right now, that is a dependency worth examining, for reasons that have nothing to do with quality.

On Sol specifically. we're in the evaluation queue and we'll be direct with clients about which work can run on it and which can't. Where access is restricted, we use the best available alternative and document why. No theatrical hand-wringing, no pretending the gap doesn't exist.

What we won't do is let the policy noise substitute for the actual work of picking the right tool for the task. That part hasn't changed.

The bigger pattern to watch

Sol is one data point in a longer trend. The US is not the only government that will try to shape frontier AI access. The EU AI Act creates its own compliance layer around high-risk use cases. China's generative AI regulations set different constraints for models deployed there. The fragmentation of the global model landscape is not a bug in anyone's plan. It is the plan.

For agencies operating internationally, the practical question is how much of your capability depends on access that a government can restrict tomorrow. That question doesn't have a clean answer yet. But asking it now, before a client's project is in flight, is better than asking it when the access denial arrives.

Frontier AI is still a creative and strategic tool first. The government has just reminded everyone that it is also, at the highest capability tier, an instrument of national interest. Both things are true. Working with that complexity, rather than flattening it into either pure optimism or pure alarm, is where the actual craft of this work lives.