Tech • AI • Robotics • Game

ENFR
TodayPlay for freeFury BloxLeaderboardShortsTop StoriesFor youTopicsVideosYT channelsArchivesSearchFavorites

Gemini 4 Carbon tops GPT-6 Astra as Alphabet drops 12.5%

AISaturday, October 31, 2026· 16 videos

Briefing

Audio player

Gemini 4 Carbon takes No. 1

Google vaulted back to the top of major model rankings with Gemini 4 Carbon, announced on October 28. The model scored 59 on the Artificial Analysis Intelligence Index, ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57, a sharp jump from Gemini 4 Argon at 50 two weeks earlier. Google also cited first-place finishes in 12 of 14 core benchmark tables, a 70.4% result on VAL, and a leading 1,528 on the blind-vote Arena text leaderboard. The launch marks Google's strongest benchmark showing in months after OpenAI and Anthropic had dominated momentum.

Alphabet selloff tests AI economics

Alphabet shares fell about 12.5% after earnings, in what was described as the company's worst trading day since its 2004 listing. Revenue of $127.6 billion slightly missed expectations, while EPS of $2.12 trailed consensus near $3.01; even excluding a SpaceX markdown worth roughly $0.55 per share, profit still disappointed. Investors focused on the cost of AI infrastructure, weaker monetization in AI-powered Search, and deteriorating free cash flow as data-center spending accelerates. The reaction crystallized a broader market question: whether frontier-model spending is outrunning near-term returns.

Big Tech enters $700B buildout

Microsoft, Google, Amazon and Meta are collectively spending roughly $700 billion on data centers and related AI infrastructure this year. Some estimates now suggest the buildout could absorb more than 3% of U.S. GDP annually through 2032, putting it in the company of railways and electrification as a historic capital cycle. Increasingly, that expansion is being financed not only through operating cash but through debt, raising the stakes if revenue lags investment. The market's recent turbulence reflects that the debate has shifted from whether AI matters to how quickly it can pay back.

OpenAI leak points to GPT-6.2 Bell

Internal tracking logs reportedly surfaced a model called GPT-6.2 Bell, nicknamed Beetlejuice, and linked to OpenAI's Doug Trio pre-training line. Rumors around the checkpoint include claims of roughly 10 trillion parameters and progress on the Navier-Stokes problem, though none of those details are confirmed. Separately, Sam Altman said OpenAI has spent about $10 million of compute on Navier-Stokes research and more than $100 million on broader scientific discovery work. The combination is fueling expectations that unreleased frontier systems are being optimized for scientific reasoning, not only chat and coding.

OpenAI expands caps and agents

OpenAI backed away from a planned reduction to its $200 plan and instead doubled Work and Codex usage for the $100 and $500 tiers. It also rolled out Dots texting in beta for U.S. pro users, enabling direct use from a phone messaging app with image uploads, location sharing and voice messages. A new model, GPT 6.1 Terra, was positioned for screen understanding, app actions and basic computer-use tasks such as simple video editing. Taken together, the moves show OpenAI pushing harder on practical agents, not just flagship model prestige.

Carbon's access stays tightly restricted

Despite its benchmark lead, Gemini 4 Carbon is not broadly available and is initially limited to a small group of cybersecurity defenders. Google is effectively using its strongest public model as a controlled deployment rather than a mass-market release, even as it touts promotional pricing of $2 per million input tokens and $10 per million output tokens. The model's standout capability appears to be long-context reasoning: around 1 million tokens across text, images, speech and video, with 96.2% on a long-context retrieval test. That restricted rollout underscores how frontier capability and safety gating are now advancing together.

DeepMind unveils Genie 4 world model

DeepMind also introduced Genie 4, a new world model aimed at simulating environments rather than only predicting text or images. The release broadens Google's narrative beyond leaderboard gains by showing continued investment in embodied and interactive AI systems. Alongside that, Mistral was described as strengthening its open-weight position, reinforcing a split market between proprietary frontier labs and more accessible alternatives. The emerging picture is of competition spreading across benchmarks, open weights, agents and simulation.

Kling 4.0 leads video test

In a side-by-side comparison of eight video generators using the same cinematic prompt, Kling 4.0 ranked first with 35 out of 40. The test used a 10-second, 16:9 prompt involving an astronaut in a rainy neon city, scored on prompt accuracy, motion and physics, consistency, and realism. Kling 4.0 was credited with the smoothest camera movement and one of the few successful renderings of a required head turn, finishing ahead of Seedance 2.5 and Gemini Omni Flash 1.1. The result adds to evidence that the generative-video race is tightening into measurable, model-versus-model competition.

Videos covered

Previous briefings · AI