Astra leaks: a 'next week' claim circulates; the only thing OpenAI confirmed is the safety brake

Leak accounts say OpenAI's largest model Astra ships next week. The confirmed core is different: per Axios, the release was deliberately slowed because its cyber capabilities approach the 'critical' threshold.

Paylaş
Astra leaks: a 'next week' claim circulates; the only thing OpenAI confirmed is the safety brake

Leak accounts on X claim OpenAI's model named Astra could ship as soon as next week. The claims are bold: Astra would be the largest model the company has ever trained, clearly ahead of the 5.6 family in early tests, with notably strong instruction following; pricing is expected around the 5.6 level. Figures circulating in the same leak clusters, such as 10 trillion parameters and a 1.5-million-token context window, also trace to the leakers, not to OpenAI. One leaker further claims Astra will leave Anthropic's Fable 5.1 behind; that comparison is an entirely unverified assertion.

What is actually confirmed

Astra's existence is real: OpenAI acknowledged the model on 1 August in a research context; there is no product, no model card, no API, no price and no date. According to an Axios exclusive on 7 August, the company has deliberately slowed Astra's release: in internal testing the model's progress in cybersecurity and coding was fast enough that OpenAI can no longer rule out it reaching the "critical" capability threshold on its own risk scale. That prompted expanded safety testing and a pause on internal activities that do not meet stricter security requirements.

On timing, prediction markets are cautious: roughly 72 percent odds of a release before December and 47 percent before November, while "this week" windows have collapsed to single digits.

In short: the model's existence and the safety brake are confirmed; the release date, the size and the rival comparisons remain rumor for now. We will update as the claims firm up.

Source: Axios · leak account

Devamını oku

Liquid AI تطلق أكبر نماذج LFM2.5 المفتوحة لتشغيل الوكلاء على الأجهزة

Liquid AI تطلق أكبر نماذج LFM2.5 المفتوحة لتشغيل الوكلاء على الأجهزة

أطلقت Liquid AI نموذج LFM2.5-24B-A2B كنموذج مفتوح، ويضم 24 مليار معامل إجمالا لكنه يشغل نحو 2.5 مليار منها في كل خطوة. وبحسب النتائج التي نشرتها الشركة، يولد النموذج المدرب للوكلاء الذين يستخدمون الأدوات 184 رمزا في الثانية على Apple M5 Max.

GlobalFeed Editor tarafından