Fable 5.1 Is Out. Claude Opened Its Memory. OpenAI Is Preparing Astra — What Is Changing in the New AI Race?

Anthropic frames the update around the architecture of large systems across several repositories, big migrations, bulk code updates and performance checks. In essence it is the same as Mythos 5.1. An ordinary chat user will probably notice no difference.

Requests to models of this class were stored for thirty days and analysed for jailbreak attempts. For a large company, data outside its perimeter means liability to clients and leak risk — and that is what held Anthropic back exactly where it aims.

With 5.1, data storage and the classifiers that decide whether to block a request can be deployed on the company's own infrastructure. The blocker is removed, and it is a move straight into the enterprise.

Repeat computation is taken from the cache, and access to it becomes about 75% cheaper. Across a whole task that is roughly 25%. For Anthropic's expensive models it is one more step towards market reach.

A "Topics" section appeared in the interface: every stored entry is visible and can be edited or deleted. It applies to Claude and Claude Cowork, but not to Claude Code.

On a long trip the model keeps the context without reminders, and the details no longer have to be explained each time. What stays open is where that context will apply later and how the model works out which place a question is about.

The Grill me skill examines an idea from every side and asks leading questions; dozens of messages are distilled into a document setting out what is being done and why. That document travels into another system and the work continues there.

A person cannot manage memory indefinitely — so the system has to: learning per project and asking clarifying questions. For now the models are solving other problems and memory stays on the user.

The expensive model in Extra High mode produced three versions of what was happening in Madrid, while a simple assistant guessed a demonstration. By evening it turned out there had been two rallies in the city. The expensive system searched and analysed, and lost the piece of information that mattered.

An ordinary model generates its reasoning token by token. In the new architecture several iterations of thinking happen inside the model before the first token appears, and what comes out is a far sparser chain.

Safety control today rests on analysing the reasoning chain: it is how an incident gets reconstructed. In hidden form that becomes far harder. OpenAI promises to limit the depth of internal thinking, but nobody guarantees the same from other companies.

Processors, high-speed memory, the network between processors, boards, server racks, software — and about nine months of development in which AI itself took an active part.

Ordinary hardware gains aggregate throughput by batching users into groups, and each of them starts waiting longer. Here the aim is to serve many requests at once and answer each of them fast.

Anthropic's opened memory will not work through someone else's interface, and Perplexity has no model of its own at the level of Fable 5.1 or Astra. The same argument explains why terminating the Cursor contract makes strategic sense.

Amazon let sellers in, watched what sold, and released its own batteries, paper and forks — they went to the top, and tens of thousands of sellers were squeezed out. The same mechanism applies to generating ads, thumbnails and short-video cuts.

The M6 and M5 Ultra have shipped, along with Mac Studio and Mac Mini, and nobody has come close to the MacBook on its own ground. But Apple has no models at ChatGPT's level — and were building them genuinely easy, Apple Intelligence would already have them.