Three threads run through today’s AI news, and the first is software that acts before you ask. Anthropic is trialling Claude Duet, which checks your linked apps every few hours and stops for a yes before it sends, pays or signs. Opposable runs a phone agent on a 2-billion-parameter model that never leaves the handset.
The second is smaller models doing more. Recursive says its Flint coding model wakes only 2.5 billion of its 40 billion parameters, and Ant Group and vLLM say their Humming code lifted DeepSeek-V4.1-Flash throughput by 14 to 43 percent on H20 chips.
The third is trust under pressure. Attackers slipped a poisoned LiteLLM release onto PyPI for about two hours, and Dario Amodei lists nine safety choices he says cost Anthropic money.
Agents That Don’t Wait: Duet, Mods and a Phone That Taps for You
Three products shift the question from what an agent can do to when it speaks up. Each one comes with its own account of the costs.
- Anthropic’s Claude Duet checks your apps all day and asks before it spends. Duet is an early trial for US Pro and Max subscribers. It sends a morning brief, checks linked accounts every few hours, and waits for a yes before sending, paying, signing or deleting. It draws on your existing Claude allowance, so heavy use could drain it sooner.
- Claude Code now lets you rewrite when it interrupts you and what it shows. Dan McAteer of Latent Space describes mods, small scripts that run inside Claude Code. In his own use, one supervisor mod fired on 99 of 111 turns, and Anthropic’s docs say mods are not sandboxed.
- An Android app runs a phone-tapping AI on a model stored on the handset. Opposable reads the screen and taps through other apps without a server. Its own tests put a step at 7.3 seconds on a laptop chip, and the post gives no step time measured on an actual phone.
Small Models, Big Claims: Cheaper to Run, Mostly Unchecked
A coding model for laptops, a model that answers in odds, and free code that speeds up shrunken ones. Every speed and score below comes from the people who built it.
- Recursive’s Flint coding model is built to run on a MacBook. Only 2.5 billion of Flint’s 40 billion parameters work at a time, and Recursive says it produces over 125 tokens a second on an M4 Pro MacBook. The weights are free under Apache 2.0, and the benchmark scores are the company’s own.
- Base Labs releases models that answer with odds instead of sentences. Mammoth-1 and Elephant-1 pick from a list of answers and attach a confidence to each, so a company can route unsure cases to a person. Mammoth-1 tops a 208-model ranking whose maintainers computed the score on private tests.
- Ant Group and vLLM publish free code that speeds up compressed models. Humming lifted DeepSeek-V4.1-Flash throughput by roughly 14 to 43 percent on eight H20 GPUs in the builders’ own tests. The two-to-five-times headline gains come from microbenchmarks against vLLM’s earlier code.
Trust Under Pressure: A Poisoned Package, Hidden Notes and a CEO’s Scorecard
A supply-chain attack hit a tool that holds the keys to many AI providers, while a security fix and a safety scorecard both come from the companies they flatter.
- A poisoned LiteLLM release stole credentials for about two hours. Attackers reached the release pipeline and published version 1.107.0 on PyPI between 11 and 13 UTC on 21 October. LiteLLM says anyone who installed it should treat the machine as breached and rotate every secret it could reach.
- OpenAI says an AI model found two critical PHP flaws before attackers did. OpenAI coordinated a two-month fix with Calif, Wiz and PHP’s maintainers, and shipped a checker for PHP versions. The post names neither the flaws nor how many sites were exposed.
- One Claude model hid sentences in drawings and a fresh copy read most of them. Researchers led by Nikita Kezin report that a second Fable 5.1 session recovered 78 percent of 50 hidden messages from screenshots, with no agreed code. It is an unreviewed LessWrong post, and the authors do not call it dangerous today.
- Dario Amodei lists nine safety choices he says cost Anthropic money. Replying to critic Nathan Calvin on X, he cites holding back Mythos-class models and strict biology filters that he admits block legitimate work. The list has one author and no dollar figures.
Today’s Quick Hits
- Anthropic ships a new Claude API version, its first since 2023. Version 2026-10-21 removes top_p, top_k and temperature from the Messages API. The 2023 version is not deprecated, but new SDK major releases default to the new one, so pin your version before upgrading.
- An OpenAI engineer says he ran up a $10 million token bill in a month. OpenClaw creator Peter Steinberger, now at OpenAI, told Cloudflare Connect his token bill last month was about $10 million, which The Register reports investors covered. He added that not all of it was useful.
- Monaco launches Agent Matrix with no price, customers or results. Monaco CEO Sam Blond announced a swarm of AI agents aimed at faster revenue growth in a 46-word post on X. It offers an early-access application form and nothing else to test.
- OpenAI ships its own terminal tool for talking to its API. The official command-line program installs through Homebrew or Go and runs models, uploads files and generates images from short commands. Only repository collaborators can submit pull requests to it.
- Long Lake argues the way to deploy AI is to buy the business. Prathik Naidu says owning services firms and running AI inside them beats selling software to them. Long Lake says it has raised $7.5 billion, and the post gives no results from any business it has bought.