Apple has secured broad access rights to Google's Gemini models. According to The Information, Apple has full access to Gemini within its own data centers and can use distillation to build smaller models from it. Gemini generates high-quality answers along with its chain of thought, which then serve as training data for a smaller model. In short, Apple is paying for what Chinese AI companies are allegedly doing in secret: tapping a powerful AI model to generate quality training data for a smaller one.
Because Apple has full access, it can build smaller versions that give the same answers as Gemini and arrive at them the same way. These lighter versions need far less processing power and can run directly on Apple devices.
Since Gemini is built for chatbots and enterprise applications, it doesn't always line up with Apple's plans for Siri, according to The Information. But Apple is still building its own models in parallel through its Apple Foundation Models team. New AI features could drop at Apple's developer conference in June.
French AI startup Mistral has releasedVoxtral TTS, its first text-to-speech model. The model supports nine languages—including German, English, French, and Spanish—and is relatively compact at four billion parameters. Mistral says it produces realistic, emotionally expressive speech and can adapt to new voices from as little as three seconds of reference audio. Latency sits at 70 milliseconds for a typical setup with a 10-second speech sample and 500 characters.
In human comparison tests, Voxtral TTS scored higher on naturalness than ElevenLabs Flash v2.5 at a similar response time. That said, ElevenLabs has since shipped a newer model with v3. Voxtral TTS is available through an API at $0.016 per 1,000 characters, can be tested in Mistral Studio, and is also available as an open-weights version on Hugging Face.
Anthropic and OpenAI are both growing fast, but they report revenue very differently, The Information reports. OpenAI's annualized revenue is around $25 billion; Anthropic's is $19 billion. Both calculate this similarly: four weeks of revenue times 13, with Anthropic adding monthly subscriptions times 12.
The key difference is how they handle cloud partners. OpenAI gives 20 percent of revenue to Microsoft and reports the number before that deduction. For Azure cloud sales, it only counts its 20 percent cut. Anthropic does the opposite: It books all cloud sales through AWS, Microsoft, and Google as its own revenue, listing the providers' shares as sales and marketing costs. Anthropic considers itself the primary provider, while OpenAI treats Microsoft as the primary provider for Azure.
Both follow US accounting rules (GAAP), but their numbers are difficult to compare. Anthropic's revenue likely looks higher on paper than it would under the same method. That matters as both companies head toward an IPO.
Google has unveiledGemini 3.1 Flash Live, its best voice and audio AI model yet. It delivers faster responses, more natural conversations, and configurable thinking levels for developers. Google says it's better at detecting pitch and emotions and more reliable in noisy environments. The model now powers live mode in the Gemini app.
According to Artificial Analysis, the model scores 95.9 percent on the Big Bench Audio Benchmark at "High" thinking, second only to Step-Audio R1.1 Realtime (97.0 percent) with a 2.98-second response time. At "Minimal," quality drops to 70.5 percent, but response time falls to 0.96 seconds.
Gemini 3.1 Flash Live scores 95.9 percent on Big Bench Audio at its highest thinking level, just behind Step-Audio R1.1 Realtime. | Image: Artificial Analysis
The model is available through the Gemini Live API, Google AI Studio, Gemini Live, and Search Live in over 200 countries. Pricing matches its Gemini 2.5 predecessor at $0.35 per hour of audio input and $1.40 per hour of audio output, making it one of the cheapest audio AI models available. The slightly better-performing Step Audio model is cheaper on input but pricier on output.
Google is making its "Search Live" feature available globally. Users in more than 200 countries can now talk to Google Search using voice and camera. Users ask questions out loud and get spoken answers with web links. With the camera on, you can point your phone at objects and ask about them—Google uses assembling a shelf as an example.
Meta is reorganizing parts of its Reality Labs division into so-called "AI-native pods" as part of a pilot program, Business Insider reports, citing an internal memo. Around 1,000 employees in the developer tools department will get new titles: "AI Builder," "AI Pod Lead," or "AI Org Lead."
The pods are small, cross-functional teams focused on delivering specific results. Engineers might take on design tasks, for example. According to the memo, the goal is a major jump in both productivity and product quality. In a statement, Meta pointed to comments from CEO Mark Zuckerberg, who said AI will change how people work in 2026 and that projects that once required large teams could eventually be handled by individuals.
Google is releasing Lyria 3 Pro, its most advanced AI model for music creation. The model can generate tracks up to three minutes long and, according to Google, has a better understanding of musical structures like intros, verses, choruses, and bridges than Lyria 3, which Google introduced in February.
Lyria 3 Pro is now available across several Google products: in the Gemini app for paying subscribers, in Google Vids for Workspace customers, on Vertex AI for businesses, and in Google AI Studio for developers. The collaborative music generation tool ProducerAI also uses the model to help create songs.
According to Google, Lyria 3 Pro doesn't imitate artists when their names appear in a prompt, but it uses them as inspiration instead. The company says the model was trained on materials "that YouTube and Google has a right to use under our terms of service, partner agreements, and applicable law," but won't share more details about the training data. All generated content is tagged with an invisible SynthID watermark.
OpenAI has reportedly finished pretraining its new AI model, codenamed "Spud," CEO Sam Altman told employees in an internal memo, according to The Information. Altman said the company expects to have a "very strong model" in "a few weeks" that can "really accelerate the economy."
"Things are moving faster than many of us expected," Altman wrote. In a related move, Fidji Simo's product organization is being renamed "AGI Deployment." To free up computing capacity for Spud and other priorities, OpenAI will shut down its video app Sora.
