🎧 Prefer to listen?

Two weeks ago, OpenAI announced GPT 5.6. It’s not generally available. The U.S. government is approving its release customer by customer, and Anthropic’s Mythos has been stuck in the same limbo for months. If you built your business on the assumption that new AI models would keep shipping on schedule, that assumption just broke.

I’ve been watching this unfold since June, when the government first pulled Anthropic’s Fable and Mythos models offline after a jailbreak was flagged as a national security risk. At the time, it felt like an Anthropic problem. Now it’s clear: it’s an industry problem. We covered the initial Fable ban and what it meant for single-model dependencies. But what’s happening now is bigger — and the implications for anyone running a business on AI tools are different than they were three weeks ago.

What actually changed

The TechCrunch report that broke the GPT 5.6 news framed it correctly: OpenAI and Anthropic are now in the same position, facing the same bottleneck, with no clear path through it. The government is testing models before release, which sounds reasonable in principle — consumer products get tested all the time. But as Dean Ball detailed in his analysis, the government doesn’t have the expertise or infrastructure to evaluate frontier AI models at the speed the industry moves. There’s no published rubric. No timeline. No clear articulation of what risks regulators are actually trying to prevent.

Mythos has been in limited preview for months with no general release date. If GPT 5.6 follows the same pattern, that’s a model that cost hundreds of millions to develop sitting behind a gate that nobody knows how to open. For the labs, that’s a financial disaster. For you, it means the steady stream of capability improvements you’ve been building around might slow to a trickle.

Why this matters for solo builders specifically

If you’re running a one-person business on AI tools, you’ve probably structured your workflows around specific model capabilities. Maybe you use Claude for long-context analysis. Maybe you rely on GPT for structured output. Maybe your automation pipeline assumes a certain level of reasoning ability that newer models were supposed to deliver.

The risk isn’t that your current tools stop working. They won’t. The risk is that the capability ceiling you were planning around gets pushed back indefinitely. If you were waiting for GPT 5.6 to handle a specific workflow — longer context, better tool use, more reliable structured output — you might be waiting a lot longer than you expected.

And there’s a subtler risk: if the approval process becomes permanent, the competitive dynamics between labs change. Right now, OpenAI and Anthropic race to ship better models because first-mover advantage matters. If every model needs government sign-off, that race slows down. The incentive to push boundaries diminishes. Models might get more conservative, more cautious, less capable at the edges where it matters.

What to do right now

1. Audit your single-model dependencies. If your entire workflow runs on one provider, you’re one government review away from a disruption. Map which parts of your business depend on which models, and identify fallbacks. We covered multi-model workflows when Fable was pulled — those strategies are even more relevant now.

2. Build for today’s models, not tomorrow’s promises. Stop waiting for the next model release to solve your workflow problems. The tools available right now — GPT 5.5, Claude Opus 4, Gemini 2.5 — are powerful enough for almost any solo builder use case. If you’re hitting limitations, the bottleneck is probably your prompt engineering or workflow design, not model capability.

3. Watch the regulatory signals, not just the product announcements. The labs will keep announcing models. Whether you can actually use them is a separate question now. Follow the regulatory developments — if the approval process gets formalized, it changes the entire calculus of which tools to invest in.

4. Diversify your AI stack. This isn’t just about having a backup. Different models have different strengths, and a multi-model approach gives you resilience against any single provider getting stuck in review. Use Zapier or Make to route tasks to whichever model is available and best suited.

5. Invest in workflow infrastructure, not model capability. The businesses that survive regulatory uncertainty are the ones built on solid automation foundations — clean data pipelines, well-structured prompts, modular workflows that can swap models without breaking. If your business only works because of one model’s specific quirks, you’re fragile.

The bigger picture

The TechCrunch analysis made a point that stuck with me: the AI industry has never had to act collectively before. It’s been a race — every lab for itself. But government regulation is a shared problem, and the solutions require trust and coordination between companies that have been trying to destroy each other. That’s a hard sell when billions of dollars are on the line.

For solo builders, the takeaway is simpler: don’t build your house on someone else’s land. The AI tools you use are products controlled by companies that are now subject to government approval processes they can’t predict or control. Build resilience into your stack. Diversify. And focus on the value you create with these tools, not the tools themselves.

The bottom line

The AI model approval process isn’t going away. Both Anthropic and OpenAI are now subject to it, and there’s no indication the government is moving fast enough to keep up with the pace of model development. For solo builders, this means treating AI model access as a variable, not a constant — and building workflows that work regardless of which models are available this week.

If you’re not sure where to start with multi-model resilience, check out the AI tool advisor for recommendations based on your specific workflow needs.