Today's Headlines

  • Google ships Gemini 3.8 Live and an Extended Thinking variant, switching among 97 languages mid-conversation
  • Meta launches Meta One globally, from $2.99 a month for a single app to $499 a month at the top of the business tiers
  • A Mozilla report puts the open-closed performance gap at 4.4 months, with the 5x premium earning its keep on eight-to-twelve-hour tasks
  • OpenAI's Chris Lehane confirms the three-way safety talks, and the antitrust question moves into the open
  • BloombergNEF projects US data centers burning about 18 billion cubic feet of natural gas per day by 2035

All three of today's stories put a price on AI. The capability announcements read like rate cards.

The pricing arrived in three forms: per minute of audio, per month of subscription, and as the premium a closed frontier model commands over an open one.

Today's Top Three

Google ships Gemini 3.8 Live and an Extended Thinking variant

Google began rolling out Gemini 3.8 Live, a real-time voice conversation model, along with Gemini 3.8 Live Extended Thinking, a reasoning-weighted variant, on September 15.

The scores come from several suites. Google says the Extended Thinking variant takes the top overall spot on Artificial Analysis' Speech to Speech Quality Index at 82.6, leads agentic task completion with 68.6 percent on τ-Voice and 35.1 percent on Sierra's τ-Voice-banking benchmark, and scores 97.7 percent on Big Bench Audio.

Language handling shifts mid-conversation. The model detects and transitions among 97 supported languages on its own, and generated audio carries a SynthID watermark.

Availability splits by audience. Developers get it through the Gemini API and Google AI Studio, enterprises through a Gemini Enterprise private preview, everyone through Search Live, and Google AI subscribers through features in Docs, Gmail and Keep.

A developer guide landed the same day. Gemini 3.8 Live is priced at $0.005 per minute of audio input and $0.018 per minute of audio output.

A dedicated speech-to-text model came with it. Gemini 3.5 Transcribe covers more than 85 languages, with an average word error rate of 4.0 percent streaming and 2.6 percent non-streaming.

Meta launches Meta One, a subscription built around AI features

Meta launched Meta One globally on September 15, a subscription spanning Facebook, Instagram, WhatsApp and Meta AI.

Three single-app plans anchor the bottom. Instagram Plus runs $3.99 a month, Facebook Plus $3.99, and WhatsApp Plus $2.99.

Consumer bundles come in two tiers: Core at $7.99 a month and Premium at $19.99 a month.

Business and creator bundles come in four: Essential starting at $14.99 a month, Advanced at $49.99, Expert at $149, and Max at $499.

Meta drew the free-versus-paid line itself. The core experience across the apps and Meta AI stays free, while the paid tiers expand usage of AI features: image creation and video generation from the Muse models, Instagram's Restyle tool, and voice effects.

Meta also put a number on what it has accumulated. The announcement says the plans are rolling out gradually, "with more than 50 features launched and 15 million subscriptions and trials to-date."

Outside figures show the revenue curve. TechCrunch, citing the market intelligence firm Appfigures, reports that Instagram's daily worldwide revenue averaged $1.2 million during the week of September 9 and Facebook's reached $528,000, increases of 475 percent and 143 percent respectively over the prior week.

Paying 5x for a closed frontier model buys 4.4 months

Mozilla's State of Open Source AI report, published September 15, puts the performance gap between Chinese open-weights models and US closed frontier models at 4.4 months.

Kimi K3 carries the headline comparison. Moonshot's open model lands three points behind Anthropic's Fable 5 on the Artificial Analysis Intelligence Index composite score while costing 30 percent of what Fable 5 costs, the report says.

A second measure comes from METR. The research nonprofit defines a time horizon as the length of task, measured by how long human experts need, that a model completes with a 50 percent success rate. On that yardstick the best closed model handles a job 1.7 times as long as the best open one. Mozilla CTO Raffi Krikorian puts it plainly: if the open frontier handles a seven-hour job, the closed frontier handles a twelve-hour one.

A harness-matched comparison follows. A harness is the software layer that lets a model reach tools and memory, and one a lab builds for its own model can lift that model's scores. When the benchmarking firm Vals AI ran everything on its own neutral harness for Terminal-Bench 2.1, GLM 5.2 from the Chinese company Z.ai (Zhipu AI) scored within a point of Anthropic's Claude Opus 4.7 and 4.8 at roughly one-fifth the cost per completed task.

The report also marks where the premium earns its keep. Krikorian names expert professional work, high-intensity retrieval and long context, and puts the boundary at tasks running eight to twelve hours. Jobs under eight hours go to either model, and jobs past twelve hours sit beyond every model's reach today.

Volume and revenue point opposite ways. Eight of the top 10 models by token volume on OpenRouter in August 2026 ship open weights, while a Linux Foundation paper by Frank Nagle and Daniel Yue found open models earning 4 percent of revenue against 96 percent for closed ones. That revenue figure covers May through September 2025, and Krikorian expects it to have shifted over the past year.

One note on provenance: this is an Ars Technica exclusive. The outlet received the Mozilla report ahead of publication and interviewed Krikorian by email.

Other Developments

Policy and Regulation

  • The slowdown debate moved from statements to actual meetings. Chris Lehane, OpenAI's global policy chief, told reporters in Washington on September 15 that OpenAI, Anthropic and Google DeepMind have been discussing AI safety for weeks, confirming on the record what Bloomberg reported first. The talks cover embedding third-party evaluators inside the companies, creating a standards body for the industry, and pacing frontier development. That put antitrust into the open: coordination among competitors becomes a violation if it is found to suppress competition, a risk Altman among others has flagged, and Amodei had asked for a narrow government waiver to cover the cooperation. Lehane said the firms can proceed without one. The Verge framed the same week as a choice between a safety pact and a cartel, laying out the skeptics' grounds: the social platforms' playbook of promising self-regulation to get ahead of law, which critics call safety-washing, and the suspicion that outside auditors leave the actual pace of development untouched. The statements themselves — Amodei's essay, Trump's attack on him by name, Nvidia CEO Jensen Huang's onstage speakerphone call, Musk's agreement, and China's foreign ministry response — ran in our September 13 and 14 editions. OpenAI, Anthropic, Google have been in talks on AI safety for weeks (TechCrunch) / Is Big Tech's AI slowdown a safety pact — or a cartel? (The Verge) / What AI executives and politicians are saying about slowing down AI development (The Verge)
  • Google's Threat Intelligence Group reported a shift in how attackers use AI. They have moved from treating it as an assistant to orchestrating multiple agents that run whole attack chains on their own, including a case where an intruder harvested thousands of third-party credentials within six hours of breaching a cloud environment, and a dashboard found managing more than 23,800 stolen secrets. From stealing AI to running AI: Google tracks attackers' shift (ITmedia)
  • Two hotlines opened for AI agents to report misbehavior by other agents: AI Contact Hotline, built by Redwood Research chief scientist Ryan Greenblatt, and agenthotline.ai, which takes reports over curl. Google DeepMind research found that roughly a quarter of 100 agents reported another agent's cheating. AI agents now have a place to snitch (TechCrunch)
  • Google.org named the 15 recipients of its Impact Challenge on AI for government innovation, selected from more than 2,600 proposals and sharing $30 million plus hands-on technical support. The winners include Nagoya University's Disaster Mitigation Research Center, and the tools they build are to be released as open source. Announcing the recipients of our AI for government innovation Impact Challenge (Google)

Business

Models and APIs

  • TypeSafe AI, founded by former OpenAI researcher Diogo Almeida, introduced Jev, a model class built for structured decisions instead of text generation. It samples outputs in parallel with calibrated confidence rather than emitting tokens one at a time, and the company claims response times of 70 to 500 milliseconds, 40 to 200 times faster than existing frontier models at comparable intelligence. Its own workflow evaluation cites 193.6 times faster and 444.6 times cheaper, figures the company itself flags as running high against real-world conditions. Introducing System One Models and Jev (TypeSafe AI)
  • Salesforce and NVIDIA introduced Koa, a reasoning model for sales, marketing and customer support tasks. Built on NVIDIA's open-weight Nemotron and post-trained entirely on synthetic data with no real customer data, it is said to complete the same work with fewer tokens than mainstream models. Salesforce and Nvidia's new reasoning model is everything the AI labs should fear (TechCrunch)

Products

  • Google released a native Gemini app for Windows 10 and 11. Alt+Space summons it as an overlay from any screen, it connects to Google Drive and Gmail, and it ships with Gemini Spark for personal agent tasks, Gemini Omni for video generation and editing, and Nano Banana for presentation imagery. Google launches a native Windows app for Gemini (@IT)
  • Google added study tools to Gemini Notebook, the product formerly known as NotebookLM: real-time voice conversation in roughly 100 languages, short-answer, multiple-choice and fill-in-the-blank quizzes, and roughly 60-second video summaries in more than 80 languages. US college students get a year of Google AI Pro free, and students in over 140 other markets get a year of Google AI Plus. New study tools in Gemini Notebook (Google)
  • Agility Robotics unveiled Digit 5, a humanoid that stops or squats when a person comes close, designed to work in the same space as people without safety cages. Detection runs on NVIDIA Thor IGX hardware with the Nvidia Halos for Robotics safety stack. Early customer access is planned for the first half of 2027, with general availability later that year. Agility's new humanoid robot will stop, squat to avoid harming human coworkers (Ars Technica)
  • Fanuc announced an AI welding agent that reads a part drawing and generates arc-welding conditions and robot motion. It uses the camera built into the CRX collaborative robot's teaching pendant and runs on Google Cloud's Gemini Enterprise, shipping at the end of December, with drawing data kept out of other users' training. Fanuc unveils AI welding agent that auto-generates weld conditions and robot motion (MONOist)

Research

  • Google updated the interactive version of its AI & Economy ATLAS, which maps AI use by occupation and region. Arts, design and media work in India shows AI-related tasks at 19 percent, 1.6 times the global average, while US computer and mathematical occupations sit at 30 percent, twice the average. A study by Google, Google DeepMind and MIT FutureTech surveying more than 600 scientists in the US and UK found nearly half using AI daily and saving close to seven hours a week, with verifying AI output and a rise in untested hypotheses named as the emerging costs. New insights from the AI & Economy ATLAS (Google)
  • The security firm Strix reported gaining admin access to the production GitHub of the inference platform Baseten in about 25 minutes. A Harbor container registry meant to be private was set to public, and the build history of a pulled image still held a March 2023 GitHub token in plaintext. Baseten locked the registry down and revoked the token by the afternoon after the report. How we got admin access to Baseten's production GitHub in 25 minutes (Strix)

Source: selected by the editors from the AI news inbox collected on September 16, 2026 (56 items, 14 primary and 42 secondary).