We may already have reached peak public AI.

Not peak intelligence. Not peak research. Not peak benchmark scores.

Peak public AI: the strongest model ordinary people and ordinary companies are legally, commercially, and institutionally allowed to use.

Anthropic's Mythos-class models are the line. On June 9, 2026, Anthropic launched Claude Fable 5 and Claude Mythos 5. Fable 5 was the broadly available, safeguarded version. Mythos 5 was the same underlying model with some safeguards lifted for trusted cybersecurity partners.

Three days later, Anthropic said the U.S. government issued an export-control directive requiring it to suspend access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including Anthropic's own foreign-national employees. Anthropic said it could not reliably verify nationality in real time, so it shut both models off for everyone.

The export controls were later lifted. Anthropic's redeployment post says Fable 5 returned globally, while Mythos 5 was restored for a set of U.S. organizations with broader trusted access still being coordinated through Project Glasswing.

But the important thing already happened.

The state found the switch.

The name almost writes itself. Mythos lives behind Glasswing, and Glasswing reveals the glass ceiling: the invisible boundary between capability that exists and capability the public is allowed to touch.

The Wall

Mythos proves there is not much point trying to sell the public a model materially beyond the Mythos line.

A lab can still build one. A government may still want one. A trusted-access program may still use one. Internal researchers may still use one to build the next generation.

But for the mass-market product tier, extra capability becomes less like a feature and more like a compliance incident.

The public ceiling is not "the best AI anyone can make."

The public ceiling is "the best AI the state will let ordinary users touch."

That changes the market. For years, the AI race was simple: build a bigger model, train it on more compute, give it tools, make it reason longer, ship it, and charge for it.

Mythos breaks that loop. At some point, the better model becomes less like a product and more like a controlled material. Distribution stops following capability. Access becomes conditional.

The Catch-Up

The rest of the industry will catch up to the wall quickly.

METR's time-horizon work estimates that the length of tasks frontier AI agents can complete has been doubling roughly every seven months. Their paper gives Claude 3.7 Sonnet a 50% task-completion horizon around 50 minutes and argues that, if the trend generalizes, AI systems could automate many software tasks that currently take a human a month within about five years.

Stanford's 2026 AI Index points the same way from the benchmark side: SWE-bench Verified performance rose from 60% to near 100% in a single year.

Anthropic's own research is even more direct. In "When AI builds itself", Anthropic says Claude Opus 4 averaged about a 3x speedup over starting code in May 2025, while Claude Mythos Preview reached about 52x by April 2026 on the same kind of optimization workflow.

None of those trends are perfect forecasts. They measure different slices of capability. But they agree on the shape: useful AI capabilities are improving on the scale of months.

So six months to a year is a plausible window for other frontier systems to reach the public-access ceiling. The point is not that every lab becomes equally good. The point is that once several labs can reach "maximum permitted AI," the public no longer experiences the race as unlimited capability growth.

The off-the-shelf models still improve.

They improve toward a public maximum.

After that, overshooting the wall is not a consumer feature. It is a reason to gate the model.

The New Race

Once the broad market is pinned near the permitted ceiling, capability becomes less differentiating.

The race becomes price, latency, reliability, interface design, memory, tools, privacy, deployment, procurement, and how often the model refuses to do useful work because the safer version had to be made safe enough to ship.

This is already visible below the frontier. Alibaba describes Qwen as a family of large language and multimodal models provided to the open-source community, with hybrid thinking modes that trade off reasoning performance, speed, and cost. Alibaba Cloud's Model Studio pricing reads like infrastructure pricing: tokens, tiers, context lengths, and model families.

That is the future shape of most AI consumption.

Not mysticism.

Metering.

Chinese models make the pressure obvious. Their appeal is not only capability. It is cost, availability, and deployability. Rest of World reported that U.S. companies are experimenting with Chinese open models like Qwen and Kimi, often through U.S. infrastructure or self-hosting to reduce data and geopolitical risk.

When the middle catches the ceiling, price pressure gets brutal.

The China Mirror

The United States is not the only country discovering that model capability is a sovereignty problem.

China has regulated public generative AI for years. The Interim Measures for Generative AI Services apply to public generative AI services in mainland China, and Stanford's DigiChina notes that providers must pass security assessment and algorithm-filing requirements before offering public services.

Now the export layer appears to be arriving too. According to reports based on Reuters, Chinese officials have discussed restricting foreign access to the country's most advanced AI models, including models from Alibaba, ByteDance, and Z.ai. The Next Web described meetings about limiting overseas access to top Chinese models, including closed and open-weight systems.

The pattern rhymes.

First, capability is soft power.

Then capability becomes leakage.

Then leakage becomes policy.

The U.S. wants to prevent advanced American models from strengthening foreign adversaries. China wants to prevent advanced Chinese models from becoming infrastructure for foreign users in ways Beijing cannot control.

Different ideology. Same control surface.

The Product After The Ceiling

Open-weight models make the ceiling porous. Once weights are public, access controls become much harder. Users can copy, quantize, fine-tune, route, distill, and deploy.

But serious institutions still need contracts, support, logs, procurement approval, indemnity, cloud access, and legal cover. For them, the practical question is not "can someone somewhere run the model?"

It is:

Can we use this model without creating a legal, security, or governance problem?

That is where the glass ceiling bites.

The lab can build something more capable than you are allowed to use.

The government can decide which partners are trusted enough.

The enterprise can decide the model is too risky for customer data.

The cloud provider can decide the export-control exposure is not worth it.

The user gets the best model that survived all of that.

For most of humanity, that model will still be astonishing. It will write code, read documents, find bugs, tutor students, draft contracts, summarize medical literature, translate across languages, automate office work, and make many mediocre software products feel haunted.

But it will not be the real frontier.

The real frontier will live behind trusted-access programs, government review, classified benchmarks, export-control lawyers, national champions, and infrastructure deals.

That is peak public AI.

Not the end of capability.

The end of naive access to capability.

Everything after that is price, usability, trust, and permission.