Today's Thread
This issue has two strands. The first is a shift for users. Anthropic cut the price of its top model, Claude Fable 5.2, by 20%; Google let AI work on its own for long stretches in the cloud through Antigravity; and LG AI Research opened K-EXAONE 2.5 at no charge. Good AI can now run cheaper and for longer.
The second is the rules attaching to places where AI stands in for people. From November 27, ads in which AI-made fake doctors or pharmacists recommend food are banned by law, and after AI assistants called Korea's National Health Insurance call center asking for proxy consultations, the counselors' union demanded the right to refuse them. As fast as you adopt tools, it is time to set standards for how AI deals with customers and institutions.
Claude Fable 5.2: 20% cheaper, output up to 30% faster
Anthropic released Claude Fable 5.2 on November 9. Developer pricing is $8 per million input tokens and $40 per million output tokens, 20% less than Fable 5.1 ($10 and $50). A token is the small unit AI uses to count text, and it is the basis for billing. Output is up to 30% faster than 5.1, and a fast mode that costs twice as much generates up to 2.5 times faster.
Fable is the general-release version of Mythos, which is open only to vetted organizations because of security risks. Mythos 5.2, released the same day, is again limited to organizations in the cybersecurity and life-sciences verification programs. During launch week, subscribers get one to three Fable usage-limit resets depending on how long they have subscribed, and Pro plans and standard Team seats can use Fable for up to 50% of their weekly limits.
Even after the cut, Fable costs twice as much as Opus 5.5 ($4 input, $20 output). Anthropic itself recommends starting most coding and document work on Opus 5.5 and moving up to Fable only for the hardest, most open-ended technical work. Keeping daily work on Opus 5.5 and sending only design or analysis problems with no clear path to Fable makes costs easier to manage.
Performance comparisons are Anthropic's own measurements. On the SWE v2 coding test, Fable 5.2 scored 69.9% and Opus 5.5 scored 62.3%. The announcement does not say when the resets expire or how long the 50% allowance lasts, so check your account settings.
Antigravity Remote Agents keep working after the laptop closes
Google added Remote Agents to its coding AI app Antigravity 2.29 and has been rolling them out since the first week of November. An agent is an AI that breaks a task into steps and carries them out on its own. Remote agents run that work not on your computer but on a Google-managed cloud virtual machine, a computer rented over the internet. Work continues even if you close the browser or put the laptop to sleep.
They are available to Google AI Pro and Ultra subscribers signed in with a personal Google account. The free plan, Google AI Plus, and business accounts such as Workspace and Google Cloud are excluded. Start from New Remote Project in the app's project menu, from the web, or with agy --remote-agent on the command line. You can attach GitHub repositories, Google Drive docs, sheets, slides and folders, and uploaded files (8 MiB per file, 50 files and 100 MiB per project).
This means you can set off a long sorting or analysis job and walk away. You can approve permission requests and check progress from your phone, so even non-developers can try attaching a set of Drive documents and asking for a summary table. Business accounts are blocked, though, so check company policy before moving work files to a personal account.
Remote agents cannot read files on your own computer; they see only what you attach. The default permission setting runs commands without asking, so change it before attaching important material. The official docs do not yet explain how usage is counted.
K-EXAONE 2.5 API opens at no charge, end date not set
On November 9, LG AI Research opened the K-EXAONE 2.5 API at no charge. An API is the connection point other programs use to call an AI model. Developers, startups, companies, and research institutions can get a key on the EXAONE API platform and use it for service development and research.
It is a technology-verification model built by further training the model released in August. LG's own scores on public data are 59.2 on the SWE-Bench Pro coding test and 78.0 on the BrowseComp web-search task, and LG says it is competitive with China's DeepSeek-V4.1-Flash and GLM-5.3. The same platform also lists EXAONE Tabular, which predicts values from table data.
It is a chance to test Korean-language work automation on a domestic model without cost. Run a few of the summarizing, sorting, or drafting jobs you now send to foreign models and compare quality and speed to see what could move.
LG's announcement does not say how long it stays free of charge, what the daily call limits are, or what the commercial-use terms are. Read the platform terms before plugging it into a customer-facing service.
Korea bans AI fake doctor and pharmacist food ads from 11/27
Korea's amended Food Labeling and Advertising Act, promulgated on May 26, takes effect on November 27. It bans ads that use AI-made virtual voices, images, or video that are hard to tell from real ones to make it look as if experts such as doctors, pharmacists, Korean-medicine doctors, veterinarians, or professors endorse or recommend a food. Violations carry up to five years in prison or a fine of up to 50 million won, and both can be imposed.
The amended text has no exception for ads labeled as AI-made or as featuring a virtual person. Separately, the Fair Trade Commission's endorsement guidelines, in force since June 1, require ads using virtual people to disclose that they are virtual. Labeling is a baseline duty, and making a food look expert-recommended is banned regardless of labeling.
Shops selling food or health foods and their ad agencies should review AI-model ads now running. Material that could be mistaken for an expert, such as an AI figure in a white coat recommending a product, is safer pulled before the law takes effect.
We could not confirm the schedule for extending this to drugs, cosmetics, and medical devices, as reports differ. The sources we checked also do not say whether the same standard applies to health functional foods.
AI assistants call Korea's NHIS; counselors want right to refuse
On November 4, an AI introducing itself as someone's work assistant called the National Health Insurance call center and asked for a proxy consultation. When the counselor said it could not be done unless the caller was the account holder, the call ended. The same day, another AI assistant asked for a basic consultation that needed no personal information and kept asking questions, and the counselor could not end the call.
NHIS guidance says proxy consultations are not allowed without identity verification, but basic consultations without personal information are. Today counselors may end a call first only in cases of sexual harassment or verbal abuse. The union branch for NHIS call-center workers issued a statement on the 5th demanding the right to refuse AI-assistant consultations and to end calls, and on the 6th sent the NHIS a letter asking for measures by the 16th. The NHIS said it is reviewing the interior ministry's complaint-handling guidelines.
AI assistants calling on people's behalf has now actually happened at a public service desk. Companies with customer phone lines should decide in advance how to verify an AI caller, how far to help, and when to hang up. Those planning to hand inquiry calls to an AI assistant should assume the other side may refuse.
Which AI service placed the calls, and what standard the NHIS will finally adopt, are not yet known.
Discord drops Electron for Rust; test build uses 290MB
Discord CTO Stanislav Vishnevskiy confirmed on X on November 4 that the desktop app is being rebuilt from Electron to Rust. Electron is a framework for building desktop apps with web-page technology. It is easy to build with, but each app carries its own browser, so it uses a lot of memory. The early Windows build he showed, Discord Phoenix, was using 290.2MB of memory while running two accounts.
The new app draws its interface with the Rust-based GPUI. On mobile, Discord is removing React Native, keeping native screens and sharing Rust code with desktop. The app will look and work the same as now; there is no release date, but user testing is to start soon.
Many messengers and work apps are built on web technology like Electron, which is a reason office PCs run out of memory quickly. As large services strip out web foundations for performance, memory use will become a criterion when choosing internal tools too. It is also worth remembering that Electron is an easy default when building desktop apps with AI.
The 290MB figure comes from a single screenshot, and Discord has not published an official announcement or performance measurements.
Try Today
If you pay for Claude, check the usage screen in settings to see whether Fable resets were granted and when they expire. Give one stuck design or analysis problem to Fable 5.2 and compare it side by side with Opus 5.5 on the same problem.
If you run food or health-food ads, list any material featuring AI-made doctors, pharmacists, professors, or other experts before November 27 and decide what to replace.
If you have a customer phone line, write a three-line standard for calls placed by AI assistants: how to verify identity, how far basic guidance goes, and when to end the call.
Get a K-EXAONE 2.5 key, run three Korean-language work prompts you use now under the same conditions, and check the commercial-use clause in the platform terms.