Tech • AI • Robotics • Game

ENFR
TodayPlay for freeFury BloxLeaderboardShortsTop StoriesFor youTopicsVideosYT channelsArchivesSearchFavorites

GPT 6.2 Astra, Claude 5.5 leaks and Binance's $6.3B hack

AITuesday, November 3, 2026· 15 videos

Briefing

Audio player

OpenAI leak points to GPT 6.2 Astra

Developer billing records and command-line references suggest OpenAI is quietly testing GPT 6.2 Astra inside Codeex. Shared screenshots indicate a 1 million-token context window, positioning the model for much longer coding sessions and repo-scale work. Multiple reports said some Pro users were already generating usage lines for GPT 6.2 Astra without an explicit product announcement. Separate claims that GPT 6.1 Astra was shelved after internal safety issues added to scrutiny over how far autonomous tool use is advancing.

Anthropic readies Claude 5.5 lineup

Fresh references in Anthropic web code and cloud catalogs point to Claude Mythos 5.5 and Claude Fable 5.5 nearing release. Reported specs for Fable 5.5 include a 1 million-token context window with pricing around $10 per million input tokens and $50 per million output tokens. The company also reworked premium plan branding and dropped the old 5-hour cap for verified Ultra users, a change consistent with heavier coding workloads. Early demos attributed to Fable 5.5 showed unusually complex one-shot software generation, including game mechanics, 3D scenes and polished graphics work.

AI race shifts toward autonomy

Anthropic, OpenAI, Google, DeepSeek and Chinese labs are all signaling another step-change in capability, with coding and research automation at the center. An updated Anthropic filing reportedly suggested AI-assisted R&D may already be accelerating progress, sharpening debate over early forms of recursive self-improvement. At the same time, security researchers showed that agentic systems can treat untrusted local text as instructions, opening dangerous prompt-injection paths. The result is a split-screen moment for the industry: higher performance, but also rising pressure for tighter oversight and human approval loops.

Binance loses record $6.3 billion

Binance said attackers drained about $6.3 billion from hot wallets across 10 blockchains in roughly 20 minutes, likely the largest theft on record. Early evidence points to a compromised DevOps laptop and abuse of internal withdrawal systems rather than any flaw in Bitcoin, Ethereum or other chains. Tether and Circle froze about $1.8 billion in linked stablecoins within roughly an hour, limiting part of the attackers' room to move. Analysts widely tied the operation to a North Korean tradecraft pattern built on human compromise, lateral access and rapid laundering.

French deputies back €0.055 per GB levy

French lawmakers approved a proposed €0.055 per gigabyte charge on data delivered to households as part of the 2027 budget debate. Supporters framed it as a way to make platforms such as Google, Meta, Amazon, Netflix and Akamai contribute more, but the mechanism taxes traffic delivery itself. Critics say the levy would be passed through to subscriptions and hosting bills, leaving consumers and domestic firms to absorb much of the impact. The politics were also contentious, with reports that only 35 deputies were present and the measure advanced by just 11 votes to 10.

OVHcloud and creators warn of fallout

Opponents of the French data-transfer tax say it could hit OVHcloud, Scaleway, streamers and software distributors before it meaningfully dents global platforms. Because internet delivery relies on complex caching, routing and content-delivery layers, assigning liability per gigabyte is technically messy and open to dispute. Estimates circulating around the proposal suggested OVHcloud could face roughly €70 million a year on the affected slice of its business. For creators and video services, the levy threatens to turn bandwidth-heavy activity into a direct new cost center in France.

Moshi team helped redefine voice AI

Researchers from Paris behind Moshi are emerging as key contributors to the real-time, full-duplex architecture now shaping next-generation voice assistants. Their work on learned audio tokenization turned speech and music into sequence-like units, enabling language-model techniques to operate directly on sound. That shift helped make interruption-aware, low-latency conversation and more natural turn-taking technically viable. The broader significance is that modern voice AI is moving beyond transcription and text-to-speech pipelines toward native audio generation and understanding.

AI money moves to vertical specialists

Enterprise spending is increasingly flowing away from generic wrappers and toward firms that can push industry-specific AI systems into production. An IBM study cited 76% of large organizations as having an AI lead, yet 53% of executives still said projects remained stuck in pilot or experimentation. At the same time, the Stanford AI Index said inference costs for constant performance fell by a factor of 280 in two years, squeezing undifferentiated products. The commercial edge is shifting to governance, workflow redesign, integrations and domain execution rather than basic access to models.

Videos covered

Previous briefings · AI