Meta launched subscription bundles pairing its social apps with extra AI usage under the Meta One brand. Meta One is a tiered subscription that packages access to its standalone app subscriptions alongside additional AI credits, available globally starting yesterday. The launch follows the recent debut of Muse, Meta's do-everything AI assistant. The bundles were in limited testing earlier this year before today's broader rollout. Pricing details by tier were not specified in the announcement, but the structure mirrors what other platforms have done to monetize AI usage above a free threshold. (The Verge)
Brief
Wednesday, September 16, 2026 · 5 min read
Google Ships Gemini 3.8 Live With Extended Thinking
Good morning. Three of the biggest AI labs are quietly talking to each other about safety while the administration publicly tells them to stop worrying about it. Meanwhile, Google shipped a new live reasoning model, Meta launched a subscription tier, and a 44-million-parameter model that fits in 19.8 MB is running faster on a laptop CPU than most cloud APIs feel. It is a strange day to be thinking about governance.
Today's reading time is 5 minutes.
INDUSTRY
Real-time voice and video AI has been a capability gap for Google since competitors began shipping persistent, low-latency conversation modes earlier this year.
Driving the news: Google DeepMind released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, two models built for real-time multimodal interaction with an added reasoning layer. The Extended Thinking variant runs a chain-of-thought process before responding, meaning it can pause mid-conversation to work through a problem rather than answering immediately. Both models are available now through the Gemini API. No pricing tiers were announced in the release post.
- Gemini 3.8 Live is positioned for real-time audio and video input, not just text.
- The Extended Thinking variant adds a visible reasoning step, similar to what other labs have shipped in non-live contexts.
Zoom in: Live, low-latency AI conversation has become a competitive pressure point in 2026, with several labs racing to make real-time reasoning feel less like a loading spinner and more like a thinking pause. Google's previous live models handled speed but not depth. Bolting extended thinking onto a live model is the attempt to close that gap without sacrificing response time.
- Extended thinking in non-live models has existed for months; applying it to streaming interaction is the new part.
- The 3.8 version number suggests this sits below the flagship Gemini line, likely optimized for latency over raw benchmark performance.
Why it matters: Developers building voice agents or real-time assistants now have a Google-native option that can reason before it speaks, which changes how you design conversation flows. Instead of routing complex queries to a separate, slower model, a single live session can handle both quick replies and harder problems. That simplifies architecture for anyone building on the Gemini API today.
Bottom line: Google just made its live models think before they talk, which is either the obvious next step or something that should have shipped six months ago, depending on how generous you are feeling.
DeepMind ↗Get this every weekday
Two minutes, 7am ET. No fluff.
Also happening
Salesforce and Nvidia released Koa, a reasoning model trained specifically for sales, marketing, and customer support. Koa is built on Nvidia's open-weight Nemotron model and fine-tuned by Salesforce for enterprise go-to-market tasks. The collaboration is notable because it bypasses the major frontier labs entirely, using an open-weight base to build a domain-specific reasoning model. That approach puts pricing and customization control with the enterprise buyer rather than with OpenAI or Anthropic. It is a direct argument that vertical fine-tuning on open weights can replace general-purpose API subscriptions for narrow, well-defined workflows. (TechCrunch)
A startup founded by an early Anthropic hire raised $40 million to insure businesses against rogue AI agent behavior. Artificial Intelligence Underwriting Company (AIUC) closed a $40 million Series A led by Ribbit Capital, with First Harmonic also participating. The founding team includes an early Anthropic hire and the former COO of METR, the AI evaluation nonprofit. The product is essentially liability coverage for autonomous agent deployments, a market that barely existed before agents started taking real actions in real systems. The bet is that enterprises will need financial backstop before they let agents touch anything consequential. (TechCrunch)
Meta's WhatsApp Business now has an MCP server so AI coding agents can handle setup and configuration tasks. The new WhatsApp Business MCP server, where MCP is a standard protocol for connecting AI agents to external tools, lets developers hand off setup, messaging template creation, testing, and troubleshooting to an AI coding agent. The server works with several major coding agents. For developers who have avoided WhatsApp Business integrations because the setup is tedious, this removes the manual configuration layer. It is a narrow but concrete example of agents doing administrative work that previously required reading documentation. (TechCrunch)
The thread
Enterprise AI skips the frontier labs
Two stories today show the same move: large enterprises building capable, task-specific AI by combining open-weight models with their own data, rather than paying per-token to a frontier API. Salesforce and Nvidia built Koa on an open-weight base specifically to own the sales and support vertical. Meta's WhatsApp MCP server hands configuration work to whatever coding agent a developer already uses, rather than locking them into a Meta-native tool. For builders, this means domain-specific performance is increasingly available without a frontier-lab subscription, but it also means the integration and fine-tuning work lands on you.
From the publisher
Snippt
Image, video and voice tools in one place
On our radar
- A developer trained a 44-million-parameter quantized model on 45 billion tokens that ships as a 19.8 MB file and runs at roughly 1,900 tokens per second on CPU, a follow-up to the SHADOW-250M model posted three weeks ago on r/MachineLearning.
- Prior Labs released TabPFN-3.5, a tabular foundation model that tops both TabArena and BeyondArena benchmarks and handles up to 1 million rows and 20,000 features, with a fast variant in alpha running 6x quicker than the base.
- A new arXiv paper called BLINDSPOT proposes a benchmark for testing whether long-horizon tool-using agents know when to refuse or pause, arguing that safety failures in agentic systems often only appear after several steps, not on the first action.
- Researchers published a position paper arguing AI is not ready for strategic wargames, citing brittleness in escalation modeling and crisis response as specific failure modes that make LLM-based military simulations unreliable.
- A separate arXiv paper found that large language models appear to represent something like pain as a distinct internal state, separate from fear or sadness, and take actions to reduce it, raising questions about model welfare that the authors say are not yet resolved.
- The AI graveyard list at TechCrunch was updated to include Apple's repeatedly delayed Siri AI overhaul and OpenAI's troubled super-app launch among projects that shut down or missed expectations in 2026.
Get the brief in your inbox
Every weekday morning. Two minutes, no fluff.