LodeHQ
AI news for people who shipFree, by email

AI · Friday, 13 November 2026 · Issue 169 · 13 stories

AI agents win support work, but United blocks them

AI is moving into real customer workflows: Airbnb’s assistant now resolves 40% of support issues and cuts cost per booking 17%, while United blocks agents from booking and check-in over safety and fraud risks. Deployment is advancing, but control remains the limit.

Models & Releases

Shopify’s shared discovery model lifts Shop feed conversion 9.85%

Shopify’s production model now generates candidates for both search and recommendations, giving Shop’s home feed a 9.85% relative lift in conversion rate. It jointly trains an open-weight language model on commerce data, using buyer behavior and product knowledge to choose what each person sees. The same model also improved search quality by 2.1 points on OneDCG@10 overall.

shopify.engineering ↗12 min read at source

Gemini 4 Hydrogen is Google’s free tier and Search model

Gemini 4 Hydrogen gives developers Google’s new lower-cost Gemini 4 option, while also powering the free tier and AI features in Google Search. It costs $1.50 per million input tokens and $7.50 per million output tokens, only 25% below Carbon and Argon today, but a 75% discount once those models double in price on December 1. Google positions Hydrogen for fast coding, tests, and everyday work alongside its deeper models.

simonwillison.net ↗2 min read at source

Gemini Decision 1 brings video and PDFs to agentic decisions

Gemini Decision 1 adds one decision API for yes/no, choice, and rubric questions across text, images, audio, video, and PDFs, giving multimodal workflows a direct way to score content. In a video test, it gave the robot question a 0.9234055387792485 probability and the dragon question 0.04976021841520312; pricing is 10 cents per million input tokens, with free output.

simonwillison.net ↗1 min read at source

Fable 5.2 cuts Anthropic’s top-model price by 20%

Fable 5.2 now ranks first on Artificial Analysis and costs $8 per million input tokens plus $40 per million output tokens, giving developers a cheaper way to try Anthropic’s current top-ranked model. The llm-anthropic extension adds Fable 5.2 and Mythos 5.2, so you can run prompts locally through Simon Willison’s LLM command-line tool. Fable 5.2 supports five reasoning levels, but its max setting used 128,000 thinking tokens and still failed to return an image in one test.

simonwillison.net ↗1 min read at source

Research

LLMs can reason through filler tokens, but serial depth still limits them

Models can solve some multi-step tasks while emitting only dots or other filler, so visible chain-of-thought is not required for every reasoning gain. The training setup makes each filler position carry a hidden reasoning step, with attention passing information between positions; it worked on three of seven tasks across Qwen3.6-35B-A3B and gpt-oss-20b. Filler tokens add parallel computation, but not unlimited serial steps.

lesswrong.com ↗27 min read at source

Products & Industry

Airbnb’s support AI cuts per-booking costs 17%

Airbnb’s customer-support AI cut cost per booking by more than 17% year over year, while resolving about 40% of support issues without a human agent in Q3. Guests will also be able to use it by phone in the U.S. as Airbnb expands the assistant to voice support. The company says customer-service NPS is at its highest level yet.

United blocks AI agents from booking and check-in workflows

United Airlines will restrict third-party AI agents on its platforms and through travel agencies, so travelers cannot safely hand the full booking and check-in workflow to a bot. The airline says shared credentials could expose data, trigger unintended purchases, change accounts, or spend frequent-flyer miles. It also says automated check-in may conflict with rules requiring a human acknowledgment about prohibited items such as explosives.

cnbc.com ↗1 min read at source

Anthropic now runs Claude across four chip families

Anthropic says Claude now trains and serves across AWS Trainium, Google TPUs, NVIDIA GPUs, and AMD GPUs, giving the company more ways to expand compute instead of depending on one supplier. The roadshow says AMD took three days to bring online with Mythos, after earlier platforms required dozens of engineers for months. Anthropic also describes “compute multipliers,” software improvements that let the same physical hardware train a model with less compute.

simonwillison.net ↗2 min read at source

Grok Bot 1.0 turns a personal assistant into a team workspace

Grok Bot 1.0 lets teams publish one shared assistant, talk to it in the app, Slack, Microsoft Teams, or meetings, and control what it learns before it changes team behavior. A Bot starts as a private draft, then publishing opens it to a team, organization, or selected people; new learnings wait in Review learnings for the owner or a Manager. Shared connectors also let teams connect apps separately instead of exposing the owner’s accounts to everyone.

flaviocopes.com ↗10 min read at source

Policy & Safety

Nvidia says chip tracking is technically feasible, despite public pushback

Nvidia told Representative Bill Foster that it could implement geotracking for advanced AI chips, giving lawmakers a technical basis for anti-smuggling rules aimed at preventing diversion to China. The company has not said so publicly: the Semiconductor Industry Association, of which Nvidia is a member, called the mandates “new, untested, and potentially infeasible.” Nvidia’s Fleet Intelligence tool already helps customers monitor chips across data centers and can indicate their country.

technologyreview.com ↗4 min read at source

OpenAI draft would let agents share user signals to avoid bot blocks

An unadopted OpenAI Internet-Draft would let agents pass a user’s network, country, masked IP range, locale, time zone, and user agent to websites, potentially reducing CAPTCHAs and bot blocks while giving the agent operator control over what gets shared. It adds six optional fields to the Forwarded header. The draft bans persistent IDs, but unsigned signals can be forged, so sites cannot reliably know who vouches for them.

joost.blog ↗6 min read at source

Tools & Open Source

MCP can cost 40 times more context than a CLI

For agents running on Claude Sonnet 5.5, choosing a CLI over a default GitHub MCP server cut one task from 59,417 input tokens to 1,421, reducing the context and cost burden without changing the result. MCP loads every tool name, description, and JSON schema up front, while a skill reveals its instructions only when needed. OpenRouter’s Tool Search reduced 46 tools’ upfront cost to under 600 tokens and halved the full task cost on Sonnet 5.5.

openrouter.ai ↗13 min read at source

OpenAI’s MCP extensions make plugins feel native inside ChatGPT

Developers can now add ChatGPT-native entry points to MCP plugins, so users can open an app from the sidebar, view supported files in custom viewers, and attach plugin resources while composing a message. The Apache-2.0 project ships TypeScript and Python packages, plus a Bits & Bolts demo that shows the integrations in practice. Its docs and changelog define the supported extensions and API changes.

github.com ↗1 min read at source

You’re caught up.

That was all 13 stories in the 13 November issue.