Tuesday, 15 September 2026

Google has released two new speech-to-speech models, Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, to general availability across the Gemini API and Google AI Studio, with Gemini Enterprise access in private preview and a rollout across its consumer surfaces [1][2]. The company describes the pair as its most advanced live dialogue models, with the Extended Thinking variant able to reason during a call rather than pausing to think between turns [1]. Both models are shipping today for developers and rolling out to Gemini Live, Workspace apps and Search Live for end users [1].

Google reports a Speech to Speech Quality Index score of 82.6, which it says places the models first overall, alongside 68.6 percent on the τ-Voice agentic task benchmark and 97.7 percent on Big Bench Audio reasoning [1]. The Sierra τ-Voice-banking test comes in at 35.1 percent [1]. The models support 97 languages and can switch language mid-conversation, and Extended Thinking can call tools in the background without breaking the flow of the call, according to the announcement [1].

For developers, the Gemini API release notes list the two new model IDs, gemini-3.8-live and gemini-3.8-live-extended-thinking, as generally available, with the second recommended when a use case needs "higher background reasoning" during live audio interactions [2]. Google has not published a public price sheet for either model in the launch post; the blog calls the rates "highly cost-effective" without naming a number, and the developer changelog directs teams to the Live API and thinking guides for setup [1][2].

Salesforce has announced Koa, a specialised reasoning model built on top of Nvidia's open-weight Nemotron family, at its Dreamforce conference [1][2]. Koa is trained specifically to work over CRM data for sales, marketing and customer-support tasks and slots into Agentforce as an alternative to the frontier models that platform already calls out to [1][2]. Jayesh Govindarajan, Salesforce's executive vice president of AI, tells TechCrunch that "one of the reasons we hadn't done this before … the challenge has always been the lack of a pre-trained base model to start with" [1].

The company says Koa is more token-efficient than Anthropic's Claude or OpenAI's ChatGPT for comparable Agentforce tasks, an efficiency claim that would translate into a lower per-call cost when running large agent volumes at Salesforce customers [1]. Salesforce says the model was trained using synthetic data mimicking customer-service environments rather than actual customer records, which the company presents as a way to build a domain model without moving live account data into the training loop [1].

Kari Ann Briski, Nvidia's vice president responsible for the Nemotron work, describes the pairing as "the trifecta of things that you need to have: sovereign AI, time to first token, efficient reasoning" [1]. TechCrunch frames Koa as Salesforce's first reasoning model built in-house rather than licensed from a frontier lab, and SiliconANGLE reports the company says Koa matches or exceeds leading general-purpose models on CRM tasks with three times fewer errors [1][2]. Salesforce has not published a public price sheet or a benchmark table for Koa alongside the Dreamforce launch [1][2].

Chris Lehane, OpenAI's head of global affairs, has publicly confirmed that OpenAI, Anthropic and Google DeepMind have been in discussions for "weeks" on shared AI safety practices, telling reporters that the three companies are exploring third-party evaluators at each lab and the creation of an AI safety standards body [1]. Lehane said the firms "don't need" a US government antitrust waiver of the kind Anthropic chief executive Dario Amodei asked for on Saturday, according to TechCrunch [1].

Sam Altman has cautioned that coordination could still put the companies at risk of violating antitrust law if it were found to suppress competition, per the same report [1]. Crypto Briefing, reporting earlier in the week, said the shared protocols under discussion cover pre-release safety reviews, standardised risk assessments and independent evaluators, and noted that the three labs already coordinate loosely through the Frontier Model Forum with Microsoft [2]. Neither Anthropic nor Google DeepMind has published a statement on the record about the talks; the confirmation is one-sided from OpenAI so far [1].

The White House position, restated by AI adviser David Sacks over the weekend and by President Donald Trump this week, is that any coordinated slowdown would advantage China and that existential-risk framings are overblown [1]. That leaves the labs discussing a voluntary safety compact that the administration overseeing them has publicly dismissed. Operators reading this should treat any resulting standards body as an industry document, not federal rulemaking [1][2].

Nvidia chief executive Jensen Huang, onstage at the All-In Summit in Los Angeles on Monday, took a live phone call from President Donald Trump and told him "You're right. We're not going to let that happen, sir," when Trump dismissed calls to slow AI development [1][2]. The remark came moments after the summit's hosts had been discussing Anthropic chief executive Dario Amodei's call to slow the pace at which AI capabilities improve, a position OpenAI's Sam Altman and Elon Musk have publicly endorsed [1].

Fox Business reports Trump used the call to characterise fears of AI takeover as "a hoax," saying "I'm telling you, it's all a hoax" and that "the robots are not going to be taking over the world" [2]. Trump also called data centres "the oil of the next 20-25 years" and said they make people and states wealthy, per Fox's account [2]. Huang's response landed to applause from the summit audience, according to TechCrunch [1].

A coordinated pause in frontier model training would fall hardest on Nvidia, whose data-centre systems supply that training [1]. Huang's endorsement of the White House line, in effect, aligns the industry's largest hardware vendor with the administration's rejection of the Amodei plan and against the position taken this week by two of Nvidia's biggest customers [1][2].

Meta has launched Meta One, a new set of consumer and creator subscriptions that combine premium features on Facebook, Instagram and WhatsApp with expanded access to the company's AI tools, effective today [1][2]. The consumer bundles are Core at 7.99 dollars a month and Premium at 19.99 dollars a month, and the creator and business plans start at Essential for 14.99 dollars and step up through Advanced at 49.99, Expert at 149 and Max at 499 dollars a month [1][2]. Single-product plans run from 2.99 dollars a month for WhatsApp Plus to 3.99 dollars each for Facebook Plus and Instagram Plus, according to Engadget's list [2].

The AI features bundled at the higher tiers include image generation via Muse Image, video generation via Muse Video, Instagram Story editing through Restyle and voice effects, per TechCrunch [1]. Business tiers add access to a Meta Business Agent on WhatsApp, Instagram and Messenger, story scheduling, account delegation and enhanced support, per Engadget [2]. Meta declined to specify per-plan usage limits, telling TechCrunch that limits can vary by "country, surface and system conditions" [1].

Meta says its subscription lineup has already reached more than fifteen million sign-ups and trials during the earlier phased rollout of Plus tiers, which it now folds into the Meta One brand [2]. For an operator or a small business already paying for Meta Verified, the practical change is a re-tiered menu with a Meta AI usage allowance attached to each rung, not a discount on what was already free [1][2].

Mozilla has published its latest State of Open Source AI report, arguing that the performance gap between US frontier models and the best open-weights models from Chinese labs has narrowed to 4.4 months, and that paying for a closed frontier model buys roughly a four-month head start at about five times the per-task cost [1]. The report highlights Moonshot AI's Kimi K3, which it says scores three points behind Anthropic's Fable 5 on the Artificial Analysis Intelligence Index while costing about thirty percent as much [1].

Raffi Krikorian, Mozilla's chief technology officer, tells Ars Technica that a closed model "earns its premium in a few places: expert professional work, high-intensity retrieval, and long context," and that Mozilla now sees the decision to pay for a closed model as "workload-specific rather than organization-specific" [1]. He gives the practical shape of the gap as: if the open frontier reliably handles a seven-hour job, the closed frontier handles a twelve-hour one, and both bands move up over time [1].

Krikorian says most open models the world runs on are Chinese today, and that Chinese labs "are running the same playbook the Americans ran with Android — give it away, but own the ecosystem" [1]. He calls for US and European institutions with a mission rather than a market — neutral foundations among them — to build a rival open coalition rather than leave the commodity layer of AI concentrated in one country [1]. DoorDash is cited as an example of a company already routing routine work to Kimi and reserving Fable for harder tasks [1].