Gemini 4 Carbon, GPT-6.1 Terra, Anthropic's S-1 Filing and More

Blog ai

Cartoon robot holding a newspaper — Gemini 4 Carbon, GPT-6.1 Terra, Anthropic S-1, LFM3, agent-desktop, Subvocal, Binance hack

Google releases Gemini 4 Carbon

Announced October 28 as Google's next frontier model, two weeks after Gemini 4 Argon.
It focuses on cybersecurity and long-running coding and agent work. Google reports 88.2% on SEC-bench Pro and 71.0% on CWE-bench.
Google says it makes about half as many unsupported claims as Argon, and checks sources and flags discrepancies before acting (internal testing).
Pricing matches Argon at $2 input / $10 output per million tokens. Access goes first to a subset of cyber defenders.

A mid-sized model released October 26, built for computer use: it understands what is on screen and takes actions across apps.
API pricing is $1 input / $5 output per million tokens, half the price of the previous Terra and GPT-6.1 Sol.
Available in ChatGPT (Plus, Pro, Enterprise, Edu) and Codex, and in the API as gpt-6.1-terra.
OpenAI reports 50.3% on Terminal-Bench 4.0 (its own evaluation). For hard programming and deep research it still recommends its larger models.

Anthropic, the maker of Claude, publicly filed its S-1 registration statement with the U.S. SEC on October 26. It plans to list on the Nasdaq under the ticker "ANT".
The number of shares and the price range have not been set yet.
The amendment filed November 2 was reported to show about $33.2 billion in revenue for the nine months through September.

Google said schools that turn off its AI tools might also lose access to Google Search, describing search and AI as deeply integrated.
After pushback, it softened the rule so Search can be enabled on its own.

Released October 27: six open-weight models of about 2B parameters (text, vision, audio, omni and more), all branched from one backbone.
They are small enough to run on phones and PCs without an internet connection, and the text model handles a 256K context.
Liquid also published notebooks for fine-tuning on a single GPU.

A Y Combinator startup making glasses that read the movement of your mouth and jaw as you speak silently, and turn it into text.
US preorders opened October 30 ($499 early bird), with shipping planned for Q1 2027.
The company claims over 200 words per minute, measured on its own prototypes with no independent testing yet.

An open-source command-line tool (Apache-2.0) that lets AI agents control desktop apps directly.
Instead of reading screenshots, it reads and operates buttons and menus through the operating system's accessibility tree, which is faster and more precise. It supports macOS, Windows and Linux.
It ships no model of its own; you pair it with a coding agent such as Claude Code or Codex.

On October 29, about $6.3 billion in assets left Binance hot wallets. Binance blamed the North Korean group TraderTraitor; North Korea denied it.
The attackers posed as recruiters and ran fake interviews. An engineer ran malicious code handed over as a test task, which opened the way into internal systems.
Withdrawals resumed November 1, and Binance said it covered the losses from its own funds so users were not affected.