The AI You're Using Isn't the Best AI Anymore

What does "tiered access" to frontier AI actually mean in practice?

Frontier labs now ship public and enterprise versions. Claude Opus 4.7 is public. Mythos Preview and enterprise Agents SDK are restricted.

For most of 2025 and early 2026, frontier AI felt egalitarian. OpenAI released GPT-4o to everyone. Anthropic made Claude available to any subscriber. That changed in April 2026, and most people didn't notice because the public release came first.

Here's what happened: On April 15, Anthropic announced Claude Opus 4.7, a major model upgrade, available to the public. The announcement sounded like a win for AI democratization. What the press didn't lead with — because Anthropic didn't publicize it — was that Opus 4.7 is not the most powerful Claude model anymore. That distinction belongs to Mythos Preview, available only to select enterprise and research partners.

The Algorithmic Bridge, an independent AI analysis newsletter, flagged this discrepancy first. The pattern is real: frontier labs are now shipping two tiers of capability, and the gap between them is substantial enough to matter.

OpenAI is doing the same. On April 15, the company announced the next evolution of its Agents SDK — new features for sandbox execution and model-native harness control. These aren't minor upgrades. They're core capabilities for building reliable autonomous AI agents. They're also enterprise-tier only. The public API doesn't get them.

What is Mythos Preview and why did Anthropic restrict it?

Mythos has recursive self-improvement capabilities. Anthropic claims safety reasons but benefits financially by locking it to enterprise partners only.

Mythos Preview is Claude's most capable version to date, with recursive self-improvement capabilities — meaning it can iteratively refine its own reasoning and outputs in ways Opus 4.7 cannot. It's the kind of model advancement that should generate excitement from AI researchers and AI-forward companies. Instead, Anthropic handed it exclusively to a small set of partners, no public waitlist, no beta signup.

The official reasoning from frontier labs typically centers on safety and responsible scaling. The argument goes: if a model has recursive self-improvement capability, you need to monitor how it's being used before releasing it broadly. Fair enough. But there's another reason frontier labs gate capabilities this way, and it's less often said aloud: restricted access creates competitive advantage for enterprise customers and strategic partners while maintaining the public narrative of being "the most accessible frontier AI company."

Anthropic, in particular, built its brand on demystifying frontier AI and avoiding the secrecy culture at OpenAI. That brand positioning — and Claude's genuine accessibility — are why many researchers and companies chose Anthropic. But brands and business models can shift. When a company has raised $8 billion (the total Anthropic has raised across funding rounds), the incentive structure around capability gatekeeping changes. Keeping your best model restricted means your enterprise customers get capabilities competitors don't. It means you can charge more because the value is unique. It means you maximize short-term revenue.

How is recursive self-improvement changing the access equation?

Only enterprise partners understand how recursive AI works. The broader research community and startups stay years behind in production experience.

Recursive self-improvement is a meaningful technical milestone. A model that can refine its own reasoning chains, reconsider its outputs, and correct errors without human intervention is qualitatively different from one that can't. It's a step toward the kind of autonomy that frontier AI researchers have been discussing — and worrying about — for years.

But here's the structural problem: if only selected partners have access to recursively self-improving AI, then only those partners get to understand how it works, what failures look like, and how to build with it safely. The broader research community, the startups building AI products, the open-source ecosystem — they're locked out. They're still building on Opus 4.7, which is excellent but not self-improving.

This creates a capability moat. Frontier labs can say they're committed to safety by restricting recursive self-improvement. Meanwhile, their enterprise partners are three to six months ahead in production experience with systems the rest of us can't touch. By the time a capability is released publicly, if it ever is, the most-prepared builders are already customers of the company that owns the model.

Who gets to use the most powerful AI — and how is that decided?

Enterprise customers with contracts get access. Everyone else doesn't. Frontier labs don't publish selection criteria or maintain public waiting lists.

As of April 20, 2026, nobody outside Anthropic's partner circle can tell you what Mythos Preview can do beyond the official description. That's the point. It's closed. The selection criteria for partner access are not published. The waiting list (if there is one) is invisible. If you want access, you likely have to call the sales team.

OpenAI is clearer about the mechanism: if you want the latest Agents SDK features, you pay for enterprise tier. You get a contract, an SLA, and a sandbox execution environment. That's honest, at least. It's a pricing model. You know what you're buying.

But it raises a question that frontier labs haven't answered publicly: If the most capable AI requires an enterprise contract to access, what does that mean for the companies and researchers who can't afford it? Right now, the answer is: you don't get to use it. You use the public tier. You're three to six months behind. You can't prototype the same capabilities your competitors are building. You can't understand the frontier because the frontier is closed to you.

What are the stakes of a two-tier AI ecosystem for society?

Capability advantage correlates with capital. The first builders with frontier access win markets while others lag months behind in understanding.

This matters for reasons beyond business fairness. Frontier AI is becoming infrastructure — the foundation that future applications are built on. If that infrastructure is tiered from day one, then capability advantage correlates with capital. The companies that can afford enterprise contracts get to explore the possibility space first. They find the edge cases, the failure modes, the novel use cases. By the time capabilities are public, the strategic advantage is already won.

It also matters for AI safety and governance. If the most powerful models are in the hands of a small set of enterprise customers and frontier labs, then the research into how those models fail is happening in private. The broader safety community isn't stress-testing recursive self-improvement. We're not publishing findings on what goes wrong. We're not building tools to detect misuse or failure modes. We're locked out.

Anthropic and OpenAI both make public commitments to responsible scaling and safety. Those commitments are real. But they're being executed in a context where the most advanced capabilities are restricted to those who sign agreements with the companies that built them. That's not an ecosystem that's designed to catch problems early or to surface systemic risks.

What should you watch for next?

Track how long restricted capabilities remain gated. If Mythos follows Opus timelines, it'll be public in October but enterprise partners build first.

Pay attention to how long it takes for restricted capabilities to become public. With Opus 3.5 to Opus 4, the timeline was about 6 months. If Mythos Preview follows the same arc, it will be public in October 2026 — but by then, every enterprise customer will have been building with it for months. The competitive advantage accrues to those who got in early.

Also watch whether frontier labs start using "safety concerns" as the reason for gating, even when safety isn't the primary driver. This isn't cynicism — it's pattern recognition. When business incentives and safety narratives align, you can't always tell them apart. The responsibility falls on independent researchers and the public to ask hard questions about what's really being gated and why.

Finally, keep an eye on open-source AI models. If Meta, Mistral, or the open-source ecosystem can ship models that are 80% as capable as enterprise-restricted versions but available to everyone, the tiering strategy collapses. That's the pressure point that matters most. It's not regulation or outcry — it's competition.