Horizon

AI Briefing

Thursday, November 12, 2026

Top stories

Google Gemini Decision 1 Launches: A Specialized Decision-Only AI ModelAI TimesVendor-claimedSingle source

Google DeepMind released Gemini Decision 1 via a new Decisions API, designed to process up to 1,000 closed-ended queries simultaneously and return results in 150–250ms at roughly 5x lower cost than Gemini 3.5 Flash-Lite, per Google's own claims. Unlike generative models, it evaluates multimodal input and returns probabilities and confidence scores optimized for routing, tool selection, and workflow branching in AI agents. This signals a broader industry shift toward purpose-built, narrow AI models for specific agentic pipeline tasks.

AI-landscape material?Does this matter?New to you?

United Airlines Bans Third-Party AI Agents from Booking FlightsCBS NewsIndependently reportedCorroborated · 2 sources

United Airlines announced a policy barring external AI assistants from booking tickets on its platform, citing privacy risks, potential for unintended purchases, and possible violations of federal aviation regulations. The airline acknowledges it has seen few actual attempts so far, but is acting preemptively amid a broader trend of enterprises restricting AI agent access. This represents an early regulatory and policy flashpoint for the agentic AI era.

Sources: CBS News · The Wall Street Journal

AI-landscape material?Does this matter?New to you?

Anthropic IPO Roadshow: Pitching a $1 Trillion ValuationETNewsSingle source

Anthropic launched its IPO roadshow with a 42-minute investor video featuring CEO Dario Amodei, targeting a price range of $70–78 per share across 825 million shares, with reported institutional demand several times the offering volume. The company is positioning itself as a trillion-dollar AI platform play ahead of its public debut. If successful, this would be one of the largest AI-sector IPOs to date, comparable in scale to SpaceX's public offering.

AI-landscape material?Does this matter?New to you?

Get the briefing free every morning

The Horizon AI Digest, by email. Unsubscribe anytime.

Google DeepMind AlphaProof Code Verifies Signal's Cryptographic Core in Four WeeksThe DecoderVendor-claimedSingle source

Google DeepMind's AlphaProof Code system, powered by Gemini Carbon agents, mathematically verified Signal's cryptographic codebase in four weeks — a process that normally takes years — writing over six million lines of proof code with roughly 1% flagged for human review, per DeepMind's own account. DeepMind says it plans to extend the approach to critical infrastructure including power grids and medical devices. If the claim holds up, this represents a landmark demonstration of AI-assisted formal verification at scale.

AI-landscape material?Does this matter?New to you?

Alibaba Previews Qwen 4 Max: 1M-Token Context, No Weights or Benchmarks ReleasedAI TimesVendor-claimedSingle source

Alibaba launched Qwen4-Max-Preview, the first model in its Qwen 4 family, featuring a claimed 1 million token context window with image, video, and tool-calling support, available via its Token Plan API until November 25. Notably, Alibaba has disclosed neither model weights, parameter count, nor any benchmark results, making independent evaluation impossible at this stage. The opaque launch contrasts with the open-weight tradition of prior Qwen releases and warrants cautious interpretation.

AI-landscape material?Does this matter?New to you?

US Corporate AI Spending Growth Slows in October Despite Record HighsThe DecoderSingle source

According to the Ramp AI Index, US corporate AI spending grew just 1.65% in October, with OpenAI token spend up 2.1% month-over-month and Anthropic at 0.5%, while 1% of companies still account for 80% of revenue for both firms. Customer support is emerging as a new growth vertical, though analysts note it may carry lower margins than the coding use cases that have driven spending to date. The deceleration raises questions about whether the enterprise AI adoption curve is plateauing at the top of the market.

AI-landscape material?Does this matter?New to you?

Google DeepMind Releases AlphaGenome 2 for Cell-Type-Specific DNA Mutation PredictionITmediaVendor-claimedSingle source

Google DeepMind released AlphaGenome 2, trained on over 100 million cell data points, claiming significantly improved precision over its predecessor in identifying disease-associated genetic mutations by cell type and state. The model is available via API for non-profit use. Per DeepMind's own claims, the improvement in precision for linking genetic variants to specific cell contexts could meaningfully accelerate genomic disease research.

AI-landscape material?Does this matter?New to you?

Nvidia CEO Jensen Huang Under Intensifying Scrutiny; Shares Drop, Republicans Move to SubpoenaAI TimesSingle source

U.S. House Republicans are preparing to subpoena Nvidia CEO Jensen Huang over alleged ties to a woman linked to Chinese intelligence, raising national security and export policy concerns. Nvidia shares fell more than 8% in after-hours trading amid mounting speculation about Huang's tenure, following earlier losses of over 9.5% across Monday and Tuesday. CoreWeave's strong earnings partially offset broader AI stock pressure, with Nvidia recovering slightly in regular trading.

AI-landscape material?Does this matter?New to you?

Emerging signals

Airlines and Retailers Begin Formally Blocking AI Agents — A Policy Frontier Is Opening

United Airlines' preemptive ban on third-party AI agents, citing privacy, unintended actions, and regulatory risk, is part of a noted broader trend alongside retailers and financial institutions restricting AI agent access. As agentic systems become capable of acting on behalf of users across the web, businesses are responding with explicit policies before large-scale adoption even occurs. This signals that AI agent access governance will become a major enterprise and regulatory battleground in the near term.

Purpose-Built Narrow AI Models for Agentic Pipelines Are Emerging as a Distinct Category

Google's Gemini Decision 1 — optimized purely for classification and routing rather than text generation — exemplifies a trend toward specialized, cost-efficient sub-models embedded within larger agentic workflows. This mirrors the broader disaggregation of LLM pipelines into specialized components for speed and cost optimization. Professionals building agentic systems should anticipate a proliferating ecosystem of such narrow models from major labs.

AI Formal Verification Is Moving from Research to Real Infrastructure

DeepMind's claim that AlphaProof Code verified Signal's cryptographic core in four weeks, with plans to extend to power grids and medical devices, suggests AI-assisted formal verification is approaching practical deployment at scale. If substantiated, this could compress timelines for safety-critical software certification across regulated industries. This is an early but potentially high-impact signal for the security, aerospace, and healthcare software sectors.

Slowing AI Spend Growth and Concentration Risk Are Becoming Visible Metrics

With 1% of companies driving 80% of revenue for OpenAI and Anthropic, and month-over-month growth rates now in the low single digits, the enterprise AI market is showing signs of concentration and potential saturation at the high end. Customer support as a new vertical may expand the customer base but compress margins. Investors and vendors alike should track whether spending breadth increases or whether the market remains structurally top-heavy.

AI Agent Infrastructure Tooling Is Accelerating: MCP, AWS Emulator, and Pomerado

A cluster of new tools — Amazon's AI-agent-written AWS Emulator for local API replication, Pomerado's YC-backed tool for turning any website into an MCP endpoint, and the evolving MCP-OpenAPI relationship — collectively signal rapid maturation of the infrastructure layer for autonomous agents. As agents increasingly become primary API consumers, tooling that bridges legacy web systems and agent protocols is becoming a distinct and fast-moving product category.

New entrants

Gemini Decision 1 model

Google DeepMind's specialized decision-only AI model, capable of processing up to 1,000 multimodal queries simultaneously and returning probability scores in 150–250ms, priced at $0.10 per million input tokens — optimized for agentic routing and workflow branching rather than text generation.

AlphaGenome 2 model

Google DeepMind's genomics model trained on 100M+ cell data points, claiming improved precision in predicting how DNA mutations affect specific cell types; available via API for non-profit use.

Qwen4-Max-Preview model

Alibaba's first Qwen 4 series model, featuring a claimed 1M-token context with multimodal input and tool calling, available on the Token Plan API; no weights, benchmarks, or parameter details disclosed.

Pomerado tool

YC F26-backed startup that converts any website into a deterministic, self-healing MCP API for AI agents, targeting legacy portals, government sites, and platforms without native APIs.

Yogurt tool

Open-source, MIT-licensed desktop app for private on-device meeting transcription, summarization, and chat — a free alternative to Granola that keeps all data local, per the developer's self-reported claims.

Biggest movers this week

Anthropiccompany
370 mentions↑83
Metacompany
44 mentions↑30
Claudemodel
166 mentions↑25
Nvidiacompany
83 mentions↑23
Claude Fable 5.2model
22 mentions↑22
Jensen Huangperson
13 mentions↑13

China & East-Asia AI

Korea AI

Japan AI

Europe (EU) AI

Regulation updates

EUPassed

Commission Implementing Regulation (EU) 2026/2487 of 10 November 2026 laying down rules for the application of Regulation (EU) 2024/1689 of the European Parliament and of the Council as regards the establishment, development, implementation, operation and supervision of AI regulatory sandboxes

Adopted

USProposed

To direct the Secretary of Commerce to promulgate regulations related to certain artificial intelligence models, and for other purposes.

Referred to the Committee on Energy and Commerce, and in addition to the Committee on Intelligence (Permanent Select), for a period to be subsequently determined by the Speaker, in each case for consideration of such provisions as fall within the jurisdiction of the committee concerned.

US StateProposed

Trade Practices; surveillance based price discrimination and surveillance based wage discrimination; prohibit

Tracked

Get the briefing free every morning

The Horizon AI Digest, by email. Unsubscribe anytime.

Something wrong on this page? Use “Report a problem” under a story, or send feedback. Who publishes Horizon: about.

This is the free daily briefing. Subscribers get the live feed, full-text search, regulation timelines, and custom alerts.

Get full access — $5/mo