Tech • AI • Robotics • Game

ENFR
TodayPlay for freeFury BloxLeaderboardShortsTop StoriesFor youTopicsVideosYT channelsArchivesSearchFavorites

Anthropic Fable 5.2 launches as Nvidia faces China probe

AIWednesday, November 11, 2026· 34 videos

Briefing

Audio player

Anthropic unveils Fable 5.2

Anthropic launched Claude Fable 5.2 and the restricted Claude Mythos 5.2, pitching them as faster frontier models for coding, reasoning, vision, robotics and long-horizon technical work. The company cut pricing to $8 per million input tokens and $40 per million output tokens, about 20% below version 5.1, while claiming output speeds up to 30% faster. Mythos 5.2 is reserved for verified cybersecurity and biomedical users, while Fable 5.2 is the public tier with stricter safeguards. The release arrives as Anthropic pushes toward an IPO and tries to justify premium spending on autonomous, tool-using systems.

Fable demos push into hardware

The most striking Fable 5.2 demonstrations moved beyond software into engineering and formal reasoning. In one test with Diode, the model reportedly designed working prototypes including a USB conference speaker, a home voice assistant and a virtual pet, handling schematics, PCB routing, enclosure design and embedded code in less than one day. In another, 1,000 agents produced a Lean formalization of Fermat’s Last Theorem in about 30 hours, far ahead of Opus 5.5 in the same setup. The performance boost underscores a shift from single prompts to expensive multi-agent production workflows with memory, verification and orchestration layers.

Google answers with Gemini Carbon

Google introduced Gemini 4 Carbon as a major step beyond Gemini 4 Argon, targeting cybersecurity and extended multi-step work. In testing inside Google Anti-Gravity, Carbon produced cleaner code, more polished visuals and stronger gameplay in projects including a Star Fox-like game and a capture-the-flag paintball demo. The model was also framed as notably stronger on writing, bug fixing and long-horizon task execution. Its rapid arrival, barely a month after Argon, highlights the accelerating release cadence at the top of the market.

OpenAI expands Codex agent stack

OpenAI broadened its developer platform with a managed Agents API built on the Codex harness, alongside deeper security tooling and app distribution features. The company said Codex Security now spans open-source CLI tools and cloud services for continuously finding, validating, prioritizing and patching vulnerabilities, citing a 0.81% false-positive rate and rollback rates below 1% for AI-generated fixes. Separately, ChatGPT’s plugin ecosystem reached 4,000 plugins, while Sites grew to 8 million hosted apps and is reportedly used by more than 70% of employees internally. Taken together, the push shows OpenAI focusing less on chat and more on managed agent infrastructure, software deployment and secure automation.

GPT-Live makes voice turnless

GPT-Live was presented as a major architectural shift in voice AI, replacing turn-based exchanges with full-duplex conversation that can listen and speak at the same time. The system separates live conversational handling from heavier backend reasoning through delegation, allowing fast responses even when more complex work is offloaded elsewhere. Developers can tune tone, accent, interruption style and conversational behavior in prompts, and even run multiple live personas in one session. The design aims to make voice interfaces feel less like sequential command systems and more like continuous human conversation.

Astra shows robot promise and risk

Tests of OpenAI GPT-6 Astra suggest general multimodal models are moving into robotics faster than expected, even without robot-specific motor training. On the Robo Dojo benchmark of 42 manipulation tasks, Astra averaged roughly 22% success, topping about 40 specialist robot-control systems, while GPT-5.5 scored under 1%. Researchers also found the model could adapt to visual distortions such as inverted or mirrored camera input, but it remained vulnerable to unsafe instructions, scene misreads and manipulated vision. The findings point to a double-edged reality: general AI can control robots surprisingly well, but brittleness and safety failures become more dangerous once models are connected to physical movement.

Nvidia scandal deepens around China

Nvidia came under intensifying scrutiny after reports linked chief executive Jensen Huang and supplier Mega Speed to a widening export-control controversy involving China. Shares fell about 8% as reports said the board had opened an internal inquiry and U.S. authorities were examining whether more than $4.8 billion in Nvidia equipment purchased between June 2024 and February 2026 was tied to suspicious channels. Additional allegations described possible interference with Commerce Department investigators and questions surrounding Alice Wong or Alice Wang, a Singapore-linked business figure associated with companies under scrutiny. The claims remain contested, but they have raised governance, counterintelligence and leadership risks for the most important chip supplier in the AI boom.

China and Europe sharpen AI race

Beyond the U.S. model war, China and Europe showed parallel moves to shape AI infrastructure and public-facing deployment. In Beijing, Xinhua debuted humanoid news anchors Xinyan and Xinyu, built by UBTech and linked to the state-vetted Xinhua Yuanxin system to constrain outputs to approved material. In France, OVHcloud said it had finished training its first large model and plans an open-source release in late November, explicitly rejecting an alliance with Mistral AI. The contrast was stark: one effort focuses on controlled media presentation, the other on sovereign model competition and open infrastructure.

Videos covered

Previous briefings · AI