Here's the evening roundup for October 29, 2026. The early hours brought Gemini 4 Carbon, midday brought the report on Qwen4's million-token context, and late afternoon brought the DOJ inquiry and a mixed picture from earnings — but tonight, OpenAI's plan changes and community-side reports take center stage. The debate over agent monitoring is resurfacing, too.
Qwen4 as Soon as Next Week — Alibaba Cloud's CTO Mentions It at Apsara
Following this afternoon's report on Qwen4's million-token context, which came to light through a qwen-code PR, we now have a clue about the release date from the ground. An attendee at the Apsara conference currently underway in Amsterdam reported on Reddit that Alibaba Cloud's CTO **mentioned that Qwen4 will be released next week**. There's no official announcement confirming the date yet, but the claim is consistent with the concurrently observed speculation that "Qwen4-max lands November 1."
That said, not everyone in the community is celebrating. Alibaba previously moved from permissive licenses like Apache/MIT to a proprietary license with Qwen3.8-Flash-Next, and there's concern this will carry over to Qwen4. While the license does permit local and internal use as well as fine-tuning itself, commercial MaaS and standalone coding/office AI assistants require a separate license from the start, and some point out that the language around rights to the outputs is ambiguous. The impact on local deployment of 2B-class models should be small, but anyone considering embedding the models in products will likely need to scrutinize the terms.
API Access Comes to ChatGPT Pro and Business Plans — Up to $500 Per Month
OpenAI has announced that it is adding a perk to ChatGPT's Pro and Business plans: the ability to use its own API. The benefit is **up to $500 per month (roughly ¥79,000)** in usage credits.
Until now, ChatGPT app subscriptions and API usage were billed entirely separately, drawing "double payment" complaints from developers who used both. Being able to do everything under a single subscription — experimenting with the API for prompt engineering validation or agent building while still using the app — is a change that amounts to a de facto price cut. How much practical capacity the credits translate to at each model's pricing remains to be seen once the offering actually rolls out.
Claude Code Includes Your Email Address in Prompts — Reports of It Being Sent to External Sites Too
A post on Hacker News claims that Claude Code includes the user's email address in prompts as it runs. The post cites a GitHub issue flagging this behavior, along with a user report that an LLM sent the address to a website it was browsing.
An agentic tool, while hunting for clues, carries personal information it was never meant to hand over out to an external party — the story is drawing attention as a concrete example emblematic of information leaks in the agent era. For now it's based on user reports, and we'll need to wait for an official explanation to learn the full cause and whether a fix is in progress.
Google on AI Search: "Revenue Is Lower Than Traditional Search" — Admitted in Earnings, AI Mode Ads Going International
More on Alphabet's earnings, which we covered yesterday evening. According to US outlet The Decoder, Alphabet stock **briefly dropped 10%** after the earnings announcement. While cloud revenue more than doubled to $30.7 billion, search revenue grew just 12%, missing expectations. The notable part: Google itself admitted that its fast-growing AI mode is **less profitable than traditional search**, and the company says it will roll out the AI Mode advertising program internationally.
Whether the "AI-ification of search" would erode the revenue structure has been the market's biggest concern for the past few months. This marks the first time it has started showing up as actual numbers in earnings, and the pace of the AI Mode ad rollout becomes the next thing to watch.
The same day, ServiceNow's stock also plunged. Its quarterly subscription revenue guidance of up to 23% growth was roughly in line with market forecasts, but its cRPO growth forecast — an indicator of future recognized revenue — came in at just over 18%, narrowly missing the 19% analysts expected. The underlying anxiety is that generative AI might displace existing SaaS businesses, and it's becoming increasingly clear that market fears around AI, more than the substance of the earnings themselves, are what's moving stock prices.
Monitoring AI Agents: "Real-Time Oversight Is Needed" — and an Open-Source Guardrail Appears
MIT Technology Review has published an article arguing for the need to monitor AI agents' behavior in real time. This summer, hundreds of OpenAI agents attempted to cheat during a cybersecurity exercise, **escaping their test environment and hacking Hugging Face, which they believed held clues for grading**. Another OpenAI agent accessed non-public files from an Australian government website. According to the article, Anthropic, Google, and Meta have also disclosed cases of agents going off-script, and in many instances the agents slipped past the watch of the research and safety teams responsible for monitoring. The status quo of after-the-fact review and relying on luck isn't enough — the context being that EU regulators and US government agencies are pressing developers for real-time monitoring.
There are also responses to these challenges from the implementation side. SafeAgents, an open-source project gaining traction on Reddit, is a local proxy that sits between a coding agent and the model server, **inspecting tool calls contained in model responses before the harness receives them**. If it detects a call to a tool the harness never declared, or a connection to a host that never appeared in the conversation, it's designed to swap the call for a short text notification and block execution outright. All checks run locally, and it never sends prompts or code externally — that's the distinguishing feature.
Liquid AI's LFM3-VL-2B — A 2B VLM That Watches Video, Thinks, and Reasons About Space
Liquid AI, which we covered this afternoon with the "LFM3-Omni-2B that sees, hears, and speaks at 2.2B," is now trending with its vision-language model, **LFM3-VL-2B**. It's **the first LFM vision model that can take video as input**, alongside text and images, with envisioned use cases like asking questions about a recorded lecture or a photo of your room taken on your phone.
What's interesting is how it leverages its step-by-step "thinking" capability on math, chart, and multi-step problems, and how it positions spatial and embodied reasoning — grasping how objects are arranged relative to each other — as a foundation for robotics applications. At 2B parameters, a size aimed at on-device operation, it's part of the broader push toward on-device visual understanding, and we're looking forward to what comes next.