Today's Thread
Today's news comes down to how much AI you use and how you split it. ChatGPT Pro at $200 a month will include less Codex usage for the same price. As more people run several AI tools at once, new ways are appearing to put those tools on one screen or to get their results on a phone.
For people who want to run AI on their own computer instead of a subscription, an experiment that gets more out of graphics cards they already own has been released. The lesson is the same either way. Rather than putting all your work on one plan from one company, measure how much your work really uses and spread it out.
ChatGPT Pro 200 cuts its Codex allowance after 10/29
According to OpenAI's help center, anyone with an active Pro 200 subscription between September 22 and 10 a.m. Pacific on September 29 keeps the previous usage allowance through October 29, 2026, as long as the subscription stays active. After that, they move to a lower included allowance while the price stays at $200 a month. Pro 200 subscriptions started after that window get the lower allowance from the start.
Codex is OpenAI's coding tool: you give instructions in plain language and the AI edits files and runs commands. OpenAI now splits Pro into three tiers at $100, $200 and $500 a month, and the fastest mode, Ultrafast, comes only with the $500 tier. Pro currently has no five-hour limit, though weekly limits may apply. The key change is that the same plan name now gives different amounts depending on when you signed up.
Teams that use Codex for automation or internal tools may see the same work stop mid-week starting October 30. Writing down what share of the weekly limit you used over the past few weeks gives you a basis for deciding whether to stay at $200, move to $500 or buy extra credits. Plus and Pro users who hit their limit can buy additional credits.
Neither OpenAI's help center nor its pricing page states how large the cut is. The multipliers circulating online come from emails that subscribers shared, so they are hard to treat as official figures. The most reliable check is your own account's usage screen after October 30.
Tutti puts Claude Code, Codex and Hermes in one room
Tutti is a tool that lets several AI agents and people share context, files and tasks in real time in one workspace. Think of an agent as an AI worker that carries out several steps on its own once you give it a goal. Tutti is open source, meaning its blueprint (source code) is public, under the Apache 2.0 license. It currently connects Claude Code, Codex and Nous Research's Hermes using the subscriptions you already have, and OpenClaw support is in development.
Until now, passing one AI's output to another meant copying and pasting by hand. In Tutti, you can pull in another agent's past conversations, files and tasks with '@' and '+', and when you write a goal it splits the work into sub-tasks you can assign to each agent. Pending approvals and running tasks appear together on one Control Center screen. Installers are available for macOS and Windows, and everyday use needs no typed commands.
If you switch between Claude and ChatGPT, this can cut down on feeding the same background to both tools again and again. But when several AIs share one job, it gets harder to track who changed what. Start small by splitting roles, for example research in Claude and cleanup in Codex.
Tutti · VM, the cloud room for multiple people and devices, is still in Early Access, and creating a room requires an invite code. Tutti says Tutti Agent costs nothing during Early Access but usage-based billing may apply later, so the pricing is not settled yet.
PlayMCP sends AI results to your KakaoTalk self-chat
Kakao's PlayMCP is an open platform that collects MCP tools, which let an AI use outside services. Think of MCP as a universal plug for AI. In November 2025 Kakao added a toolbox feature so users can pick tools and use them in outside AI services such as ChatGPT and Claude with a single Kakao account sign-in. Kakao's own tools include KakaoTalk's chat-with-yourself room, Talk Calendar, Kakao Map, Gift and Melon.
In Claude you add the toolbox as a custom connector in settings, and in ChatGPT you register it as a custom MCP server address in developer mode. A PlayMCP connector made by Kakao was added to Claude's connector list in January 2026 and requires sign-in. As in Kakao's own example, asking the AI to send what it just told you to your KakaoTalk self-chat delivers the result to your KakaoTalk. In May 2026 Kakao also added support for OpenClaw, an open-source agent you install on your own computer, which connects using a one-time code valid for only 10 minutes.
If you often hand AI a long task and step away, this gives you a channel to get the finished result or a summary on your phone. The biggest practical difference is that you use a path Kakao opened officially, without installing another app or driving KakaoTalk with unofficial programs.
Any tool you put in the toolbox becomes usable by the AI you signed in from, so add only what you need. The OpenClaw connection is removed under Profile, Settings, Connected services in PlayMCP. Sending to group chats with other people was not confirmed in Kakao's official materials.
gpu-unlock lifts RTX 3090 AI compute by 25–67%
gpu-unlock, an open-source tool on GitHub, doubles the instruction rate for certain AI calculations on NVIDIA RTX 30-series graphics cards, which had been held below the workstation cards built on the same chips. In the author's measurements, an RTX 3090 Ti went from 87.50 to 157.41 TFLOPS in BF16, a number format common in AI. TFLOPS counts how many calculations a chip handles per second. Reported gains are 25–67% for the RTX 3090, 68–80% for the 3090 Ti and about 80% for the 3060 12GB.
The tool is a Linux kernel module, a component that sits deep inside the operating system, and it runs before the NVIDIA driver at startup, so the driver itself stays unmodified. Target chips include GA102, GA104 and GA106, and the license is GPL-2.0. The author notes that each report used its own benchmark, so the numbers are not directly comparable, and the lowest 25% came from cards capped at 225 W.
For teams hoping to run AI in the office on a used RTX 3090, this means more work from the same hardware. The felt difference depends on the job, though. Some reports showed 14–36% faster reading of prompts and documents, while the stage that writes the answer word by word stayed about the same. Gains are largest for reading-heavy work such as summarizing long documents, and short chat replies will not speed up as much as hoped.
This is an experimental tool published by an individual, not a feature the manufacturer supports. Waking from sleep adds about a second to re-apply the unlock, hibernation is not supported, and a card driving a display goes blank during the reset. It runs on Linux only and targets RTX 30-series chips only.
Try Today
If you use ChatGPT Pro 200, note this week's usage share on the Codex usage screen today and compare it with the same weekday after October 30. If the gap is large, work out then whether a higher tier or extra credits costs less.
Connect Kakao PlayMCP from Claude's connector list and put only KakaoTalk's chat-with-yourself room in the toolbox. Ask the AI to send today's notes to your self-chat and check that the KakaoTalk message actually arrives.
If you use both Claude and ChatGPT, list the background details you keep feeding to both before installing Tutti. The longer that list, the more you gain from one shared workspace.
If you run AI on an RTX 30-series card, measure your current speed before installing gpu-unlock. Record the time to the first word of the answer and the time to the end of the answer separately for a long document, so you can compare the effect fairly.