Today's thread
Today's AI news is less about performance than about price and bundling. Google launched Gemini 4 Argon with API prices at half the list rate until the end of November, and xAI folded Grok, the coding tool Cursor and X Premium into a single plan. OpenAI opened the Decisions API in public beta, which returns short judgments such as classifications cheaply and quickly. Which pricing structure you are locked into now matters as much as which model is smarter.
Platforms and regulators kept drawing lines. Meta blocked ads from TikTok parent ByteDance in seven countries, and Kakao's Brand Message is being investigated by Korea's Personal Information Protection Commission over the scope of consent. Korea's Ministry of Intellectual Property is using AI to find online counterfeits, and the question of verifying Tesla's FSD has moved to a National Assembly audit on the 23rd. When choosing marketing channels and tools, ask whether you may before asking whether you can.
Gemini 4 Argon goes GA — API at half price until November
Google made its new top model, Gemini 4 Argon, generally available worldwide on October 13. Developers and enterprises can use it through the API in AI Studio and the Gemini platform; an API is the connection point other programs use to call a model. Until November 30 it costs $2 per million input tokens and $10 per million output tokens, doubling to $4 and $20 from December. A token is the small chunk of text a model reads and writes, roughly three-quarters of an English word.
Subscribers can use it right away too. Google AI Ultra and Pro subscribers get it in the Gemini app and AI Mode in Search, and Ultra subscribers also get it in Antigravity, Google's agent development environment. An agent is an AI that chains several steps on its own to finish a task. Input and output together can reach one million tokens per exchange, and cached input, a repeated opening sent again, is 95% off.
November 30 is effectively the trial window. Compare results against your current model on the same work documents and questions, and calculate costs at the December list price, not the discount. Subscribers can try it in the app first and decide on the API later.
Most of the results Google highlights come from its own or partner evaluations. Independent comparisons of Korean-language work quality are still scarce. Whether it is worth the full price is something to judge by using it.
The X Subscription — Grok, Cursor and X in one plan
xAI launched The X Subscription, which bundles Grok, Grok Bot, the Grok models inside the coding tool Cursor, and X Premium into one plan. The official price list is Free at $0, Plus at $8 a month, Premium $30, Super $100 and Ultra $200. Included usage is 4x Plus for Premium, 15x for Super and 50x for Ultra, and Grok Bot starts at Premium.
Until now Grok, X Premium and Cursor were billed separately. Now using Grok models in the Grok app or Cursor draws from one weekly usage pool per person. Using other companies' models in Cursor, however, is not covered and is charged from the first use as additional credits. Annual billing saves about two months, for example $300 a year for Premium.
Teams already paying for Cursor and X separately should compare the combined cost. If you mostly use other companies' models in Cursor, though, the bundle helps less. Team seats from Premium up cost the same per seat as the individual price.
The price list does not state launch regions or payment terms for Korea, and listed prices exclude tax. How many requests each usage multiple actually buys has not been disclosed. Adjusting the tier after watching the first month of usage is the safer path.
OpenAI Decisions API — a classification endpoint in public beta
OpenAI opened the Decisions API in public beta. It returns only judgments such as true or false, or a pick from options you define. The only model is gpt-6-luna, input costs $0.10 per million tokens, and there is no output charge. Results come in three forms: the probability a condition holds, a choice with per-option probabilities, and a score on a scale. OpenAI says it is about 10 times faster than the regular Responses API.
Until now, jobs like sorting inquiries or routing them to a team meant giving a general model long instructions and then interpreting its answer. With the Decisions API you define the question and options up front and get an answer with probabilities attached. The official docs expect general availability within weeks.
It fits work that repeats the same judgment at volume, such as classifying customer inquiries, prioritizing email or reviewing content. Because probabilities come back too, you can draw a threshold, say auto-handle above 0.8 and send the rest to a person, to split work with human review.
The docs say you must set that threshold yourself using labeled examples. Images must be embedded in the request rather than passed as web addresses, and multi-step judgments require separate requests. There is no independent performance comparison yet.
REA plugs reverse engineering into AI coding tools
REA, a reverse-engineering tool, is public on GitHub under the MIT license. Reverse engineering means taking apart a finished program without its source code to learn how it works. REA is an MCP server that lets AI coding tools such as Claude Code, Codex, Cursor and Gemini CLI drive existing analysis tools like Ghidra and IDA. MCP is a common plug standard for connecting outside tools to an AI, and setup is one line: npx rea-agents setup.
Until now each analysis tool had to be learned separately. REA has the AI pick and run tools against desktop apps, .NET programs, Android apps and websites, and report the evidence and limits behind each conclusion. The analysis runs on your machine, but the results go to whichever AI model provider you connect, under that provider's data policy.
It can help you understand an undocumented in-house legacy app or build features compatible with an existing product. Using it to copy someone else's app raises terms-of-service and copyright problems. REA itself says to use it only for lawful research, education, interoperability, and analysis of software you own or are authorized to examine.
Setup shows the proposed changes and asks for approval before touching your configuration, and backs up existing settings. Paid analysis tools such as IDA and Hopper need their own licenses. Because lookalike repositories are circulating, the project warns to install only from its official GitHub repository and npm package.
Meta blocks TikTok ads in 7 countries — Korea not included
Meta has blocked ByteDance ads on Facebook, Instagram and its other apps in seven countries: the US, Canada, Egypt, Indonesia, Japan, Thailand and Vietnam. The ban covers not only ads ByteDance runs itself but also campaigns by other advertisers that send people to TikTok or other ByteDance services. A Meta spokesperson said the company has no obligation to run ads from a competitor trying to pull people off its apps.
The move follows TikTok's September restrictions on links that send users out to other social platforms. The two companies have also clashed over the ad market and youth safety. Korea is not among the seven countries.
Brands running Meta ads in those countries that drive traffic to TikTok accounts or TikTok Shop may see campaigns stop. Even Korean advertisers targeting Japan, Southeast Asia or North America should check, for example by pointing landing pages to their own sites.
Meta has not said whether it will expand to other countries, and TikTok and ByteDance have not yet responded.
Kakao Brand Message faces privacy probe over consent scope
Korea's Personal Information Protection Commission ran an on-site inspection of Kakao in late September over Brand Message, which lets businesses send KakaoTalk ads to people who have not added their channel as a friend. The issue is how ad recipients are identified. According to data submitted to the National Assembly, Brand Message spam reports reached 94,947 from January to July this year, about 2.5 times the 37,169 from June to December last year. Related fines rose from 2 cases last year to 22 this year.
The older Friend Talk reached only channel friends. Brand Message relies on marketing consent advertisers already hold to reach non-friends as well. A civic group filed a complaint arguing that using phone numbers collected for messaging to identify ad targets is use beyond the original purpose, while Kakao says the service is based on recipients' explicit prior consent.
Marketers using Brand Message should check whether their consent wording actually covers KakaoTalk ads. Organizing records of how consent was collected and how opt-outs were handled will speed up any response, whatever the probe concludes.
The commission has not yet announced its findings or a timeline.
Korea's IP office hunts counterfeits with AI — 35 brands first
Korea's Ministry of Intellectual Property said on October 11 that it will run joint public-private monitoring of counterfeit goods with trademark holders and online platforms through year-end. The first round, October 12 to 30, covers 35 brands in cosmetics, health supplements, children's products and cleaning products across 15 platforms including open marketplaces and social media. From November it expands to clothing, fashion accessories and shoes, which peak with cross-border shopping around Black Friday and Boxing Day.
AI flags suspected counterfeit listings; when needed, staff buy the product, the trademark holder checks authenticity, and confirmed fakes are blocked from sale. Findings go to the Korea Customs Service and the Korea Communications Standards Commission to block imports and sales sites. The program reflects a view that counterfeiting has spread from luxury goods to everyday items.
Brands selling everyday goods online should organize their trademark registrations and authenticity criteria so they can answer appraisal requests quickly. Sellers should keep purchase records so goods of unclear origin do not get caught.
The list of 35 brands and how the AI detection works have not been disclosed.
Tesla FSD test-car proposal reaches lawmakers — reply due by Oct 23
Tesla Korea's government-relations officer submitted a document proposing prior consultation with the transport ministry on supervised FSD to the office of Rep. Jang Jong-tae of the National Assembly transport committee, according to an October 13 report. The document offers safety-verification cooperation, test vehicles, and limited real-road testing of China-made Model 3 and Model Y cars under temporary operation permits. Supervised FSD is Tesla's driver-assist feature in which the car drives itself while the driver keeps watching.
At the October 8 parliamentary audit, Tesla Korea's CEO said the company had never asked the government for testing, and the ministry said there is no international standard for FSD. The document shows the company did propose verification cooperation in writing. The ministry agreed to answer on FSD rule changes before the committee's comprehensive audit on the 23rd but had not given the office its position as of the report.
The direction for FSD in Korea may become clearer around the 23rd. Companies using FSD in work vehicles should keep travel and vehicle rules on the assumption that the driver bears accident responsibility until then.
What verification process the ministry will propose, and when China-made cars might be covered, remain undecided.
Try today
Run three of your usual work documents through Gemini 4 Argon and your current model, compare the results, and calculate monthly cost at the December list price ($4 per million input tokens, $20 per million output tokens).
Pick one recurring classification task and write down the question and options. Fixing options, such as whether an inquiry is about a refund, delivery or something else, speeds up automation design whether you use the Decisions API or another tool.
If you send Kakao Brand Messages or text-message ads, pull up your consent wording and the date and channel records for each consent, and check that receiving ads via KakaoTalk is stated explicitly.
In Meta Ads Manager, look for campaigns targeting Japan, Southeast Asia or North America that send people to TikTok, and change their landing pages if you find any.