Four frontier AI models shipped in 72 hours and three came with a locked tier

Four frontier models shipped in seventy two hours. Three of them came with a version you are not allowed to use.

Anthropic went first, releasing Claude Fable 5.1 and its trusted-access twin Mythos 5.1 on 1 September, alongside a 75 percent cut to cache-read pricing. Google DeepMind followed on 2 September with Gemini 3.8 Flash and a defenders-only Cyber variant. Meta put out Muse Spark 1.3 the same day at roughly 0.10 dollars per million tokens blended. OpenAI released GPT-6 Astra on 3 September, the first model the company says triggered its critical-cyber safeguard threshold.

The best AI model September 2026 comparison is no longer one list

Benchmarks are the easy story here and the least useful one. What the four releases share is structural. Capability is being sold in lanes now, and a restricted cyber tier appeared in three of the four launches, which means the model you can call with an API key and the model that exists are increasingly different products.

That has a practical consequence for anyone building on these APIs. Choosing a model is no longer a matter of reading a scorecard and taking the top row. You have to work out which tier your organisation qualifies for, what the approval process involves, and how long it takes. None of that appears in a benchmark table.

Pricing moved twice inside one month

The second pattern is volatility. Anthropic cut cache reads by 75 percent. Meta launched at roughly a tenth of a dollar per million tokens blended. Promotional rates, cancellations and scheduled increases have all landed inside the same four week window.

Eighteen months ago, per-token pricing was close to a fixed input in a cost model. It is now a variable that moves on a quarterly cycle, and a workload priced in July may not be priced the same way in October. If you run anything at volume, the cache-read line is worth checking specifically, because that is where the largest single change of the month happened.

This is downstream of infrastructure economics rather than generosity. The electricity and capital costs behind these deployments still have to be recovered, and providers competing hard on headline token price tend to recover it somewhere else on the bill.

Access narrows at the top while distribution widens at the bottom

Gemini crossing one billion monthly users puts the scale of the general models beyond argument. The restricted variants released in the same week are going to a vetted list. Those two numbers are moving in opposite directions inside the same product families.

Capability at the frontier is still climbing as well. Astra's earlier work on open mathematical problems suggested the ceiling had not settled, and a critical-cyber safeguard trigger three days into September says the same thing in a less pleasant way.

What to watch

Whether the restricted tier becomes standard practice or stays an exception. If the fourth lab ships one in October, the industry has adopted a two-model release convention without anyone announcing it. The part worth sitting with is that this happened across a single long weekend, and nobody outside the labs had a say in it.