AI news today: Genie 4, Claude's Exponential Time Hypothesis claim

AI news today, October 29, 2026: here is what was announced on Oct. 26 and 27. Google DeepMind introduced Genie 4, a world model (an AI model that simulates an interactive world you can steer in real time).

Anthropic says its Claude model refuted the Exponential Time Hypothesis, a computer science conjecture, though that result is not peer-reviewed yet. Anthropic also details five exploit chains (bugs combined into one working attack) and a 40% government discount. Axiom, Artificial Analysis, Apple and Cloudflare each have news too.

Google DeepMind announces Genie 4, a real-time world model with audio

DeepMind says Genie 4 generates interactive, photorealistic worlds in real time, and that it is the first of its real-time world models to generate audio natively, including ambient sound, music and dialogue. Users steer a scene with text prompts, for example to change the weather, move the camera or add a person. DeepMind says the model builds on the Gemini family, with Gemini Omni anchoring the simulation in real-world appearance and behavior (DeepMind).

Genie 4 is a research preview for trusted testers. DeepMind says it is coming soon to Google AI Ultra subscribers in Project Genie and to early-access partners of the Gemini API and Google AI Studio. DeepMind also says visual details and object permanence can drift in longer sessions.

Why it matters: DeepMind frames world models as a way to build training environments for robots and agents, and as a base for new interactive experiences. It is still an early research preview, not a public release.

Anthropic says Claude refuted the Exponential Time Hypothesis (ETH)

Anthropic says an internal Claude research model found a deterministic algorithm that solves 3-SAT (a classic logic problem) in time 2^o(n), which would refute the Exponential Time Hypothesis, or ETH, the belief that 3-SAT needs exponential time. Anthropic’s paper, “A 2^o(n) Time Algorithm for 3-SAT,” is credited to Claude and was posted as a preprint on Hexagon, a math preprint site. It was posted on Oct. 27 and has not been peer-reviewed.

Anthropic says the core algorithm was found on October 16 with 14 million output tokens (chunks of text the model generates) and no human input, and that its researchers then worked with Claude on the presentation. The paper also notes that an internal OpenAI model “independently and concurrently” claims a similar result, which we covered in yesterday’s math roundup.

Anthropic’s GitHub repository includes a formalization in Lean (software that checks each step of a proof). Its README says AI models wrote the Lean code and that no machine checks that the main statement file says what the paper says. It also says the result is purely asymptotic and says nothing about solving formulas in practice.

Why it matters: Two labs reporting similar results makes them harder to dismiss, but both are still claims. Independent expert review is the next step.

Axiom says AI agents formalized the finite simple groups proof in Lean

Axiom says it has released the first complete computer-checked proof of the Classification of Finite Simple Groups, whose original proof spans roughly 15,000 pages. Axiom says its mathematicians directed swarms of dozens of AI agents for 24 days, producing about 43.1 million lines of Lean. It says the process exposed 34 noteworthy deviations from the published record and that it incorporates about 1.2 million lines of code from the existing Qiuzhen-CFSG project (Axiom; repository).

Asked whether anyone outside Axiom had rebuilt the proof, CEO Carina Hong told RuntimeWire by email, “Not yet. We are publishing the build artifacts to help and encourage external parties to independently reproduce it.”

Why it matters: If it holds up, it is a large demonstration of AI-driven formal math. Like OpenAI’s earlier claims (see our earlier roundup), it is the company’s own claim until others reproduce it.

Anthropic’s red team details five exploit chains found by an unreleased Claude model

Anthropic’s Frontier Red Team says an unreleased internal research model found critical vulnerabilities in widely used software, and it describes five attack chains that led to remote code execution (an attacker running their own code on someone else’s machine) and other critical impacts with little or no user interaction. They cover Chrome, Safari, an Android lock screen bypass, a Windows denial of service and a zero-click Linux kernel exploit (Anthropic).

Anthropic says it reported each issue to the affected vendor under a coordinated disclosure policy that gives developers 90 days to patch, and that fixes are available for the cases it describes.

Why it matters: Anthropic says models can now chain multiple bugs into working exploits with little human involvement. Its advice to defenders is to scan their own code.

Anthropic and GSA cut Claude prices 40% for government buyers

Anthropic says an expanded agreement with the U.S. General Services Administration (GSA) gives government buyers a 40% discount on Claude Enterprise, Claude for Government and the Claude API, with no minimum spend. The discount also extends to state, local, tribal and territorial governments (Anthropic).

Anthropic estimates the discounts will save the government $1 billion and says about 500 U.S. government organizations now use Claude. GSA’s release describes it as a 40% discount on token-based usage (pay-per-use pricing) with no minimum orders (GSA).

For more on Claude and its rivals, see our AI assistants compared list. Anthropic’s IPO filing is in our earlier roundup.

Why it matters: GSA says OneGov is moving toward consumption-based pricing, and Anthropic says agencies are not bound by a minimum spend. That lets agencies start small and pay only for what they use.

Artificial Analysis launches Intelligence Index v5

Artificial Analysis says v5 adds two new benchmarks, FrontierCode 2.0 and Terminal-Bench-Science, and puts 60% of the index on private or held-out test sets. It reports Claude Opus 5.5 (max) leading at 58, ahead of GPT-6 Astra at 57 and GPT-6.1 Sol at 55, with Gemini 4 Argon (high) at 50. It says the gap to open-weight models (models whose trained parameters are public) widened from 3 points to 12, with Kimi K3.1 Pro, whose weights are not yet released, the top open-weight model at 46 (Artificial Analysis).

Why it matters: Artificial Analysis says GPT-6.1 Sol scores within 4 points of Claude Opus 5.5 at about an eighth of the cost per task, so price matters as much as rank.

Apple adds AI assistants and MCP support to Keynote, Pages and Numbers

Apple says Keynote, Pages and Numbers get an in-app assistant (beta) built on Apple models and OpenAI. They also get Model Context Protocol (MCP) support, which lets external AI agents read content, make edits and run automated actions. Keynote also adds an open JSON-based file format, .keyx (Apple).

The updates arrive on the App Store in November as a free update for existing Apple Creator Studio subscribers. The assistant needs an Apple Intelligence-capable device running iOS 27, iPadOS 27 or macOS 27 or later, and is U.S. English only. MCP requires macOS 27 or later.

Why it matters: MCP is a standard way for AI agents to connect to apps, and Apple is adding it to its own productivity apps.

Cloudflare discloses a Workers AI data exposure tied to an SGLang bug

Cloudflare says a configuration mismatch and a bug in the open-source SGLang inference engine (the software that serves AI models) caused 155 responses from its GLM-5.3 Workers AI model to contain data from other customers’ requests. It says this happened between August 27 and September 9, only for prompts longer than 681,280 tokens, and that it found no evidence of deliberate exploitation.

Cloudflare says it fixed its configuration and the engine, and advises SGLang operators to upgrade to v0.5.22 or later (Cloudflare). SGLang’s own security update says a related cache-release bug is fixed by default from v0.5.21 for Mooncake and NIXL setups, and from v0.5.22 for MORI.

Why it matters: It is a detailed public account of an inference-serving bug that exposed data between customers. Anyone self-hosting with SGLang should check their version and that the KV cache (the model’s working memory for a prompt) can hold the full context length.

Also notable

Liquid AI released LFM3. Liquid says it is six open-weight models of about 2 billion parameters (text, vision, audio, omni and two decision models) on one backbone, and that the text model fits in about 1 GB at 4-bit precision (Liquid AI).
OpenAI and Oracle expanded their partnership. OpenAI says Oracle is developing Oracle AI Workspace to bring Oracle applications and data into ChatGPT, and that U.S. Oracle customers can buy OpenAI API, Codex and ChatGPT Work through Oracle Marketplace. The post gives no launch date for Oracle AI Workspace (OpenAI).

Next edition: AI news today: Google’s Gemini 4 Carbon, Mistral Large 4.