Skip to content
AI 24 UpdatesAI 24UpdatesWhat changed. Why it matters to you. What to do.

ChatGPT for Healthcare computer-use agents and plugins; 66 regional voices in the Gemini app; Gemini Notebook interactive sites; Claude Code iPhone Duo simulator; Liquid AI d1.2 decision models; Tencent Hy4 flash; Cohere Video Intelligence in Compass

Version 1

Share on X

In Brief

  1. OpenAI — ChatGPT for Healthcare adds computer-use agents
  2. Google — 66 regional voices roll out in the Gemini app
  3. Google — Gemini Notebook turns sources into interactive sites
  4. Anthropic — Claude Code desktop adds iPhone Duo in the iOS simulator
  5. Liquid AI — d1.2 and d1.2-ultra omnimodal decision models
  6. Tencent — Hy4 flash released
  7. Cohere — Video Intelligence in Compass enters private preview
Abstract editorial illustration of computer-use agents moving information between healthcare system panels, with the caption OpenAI — ChatGPT for Healthcare agents

Details

OpenAI — ChatGPT for Healthcare adds computer-use agents

OpenAI

ChatGPT for Healthcare adds computer-use agents, stronger clinical search, and more than 20 new healthcare plugins.

  • Computer-use agents, improved trusted clinical search, and more than 20 new healthcare plugins are available in ChatGPT for Healthcare; computer-use agents and the improved clinical search are also available in ChatGPT for Clinicians and, for computer use, in the OpenAI API.
  • The agents carry out multi-step work in existing healthcare software by reading screens, clicking, typing, and moving between applications. Organizations can set which applications agents may use, which steps need human review, and can keep audit logs. OpenAI says content shared with ChatGPT for Healthcare is not used to train models, and that HIPAA-compliant workflows involving protected health information require eligible services covered by a Business Associate Agreement plus the required security and retention settings.
  • OpenAI's results on EHR inbox tasks with GPT-6 Astra: 99% of 325 reviewed tasks were rated safe, and physicians selected “very helpful” or “extremely helpful” in 93% of 646 helpfulness assessments. On HealthAdminBench, an independent Stanford benchmark of administrative workflows, GPT-6 Astra earned 94.3% of available rubric points, while GPT-6 Luna earned 91.8% at 8.5 cents per task, 98% less; both were evaluated at maximum reasoning effort.
  • OpenAI says it is working with Abridge, Ambience, Cigna Healthcare, and CVS Health on computer-use applications such as referrals, document processing, and gathering information across systems. CVS Health is open-sourcing its Agent Kit, built with OpenAI's Agents SDK.
  • In a blind evaluation of the updated clinical search, physicians reviewed responses to real-world clinical questions, including multi-turn conversations. In more than 1,300 reviews, OpenAI says physicians rated the updated experience more highly across all eight dimensions, including factual correctness, safety, actionability, and source authoritativeness.
  • The new plugins, available in ChatGPT for Healthcare including through ChartGPT, include Aiva, Amion, Assort Health, Availity Extend, Elation Health, Ferry, IMO Health MCP Server, Infinitus Relay, K Health, Luma Health, Metriport, Notable Integration Specs, Oracle HDI Overview, PenguinAI Medical Coding, Photon, Redox, Stedi, Subsalt, Trizetto, Turquoise, and UpToDate Expert AI. ChatGPT for Clinicians is free for eligible, verified clinicians in the United States.
Sources:OpenAI (computer-use agents)✓ officialOpenAI (ChatGPT for Healthcare)✓ officialChatGPT release notes✓ officialOpenAI on X✓ official

Google — 66 regional voices roll out in the Gemini app

Google

The Gemini app is rolling out 66 new regional voices on Android and iOS, including several English, Hindi, Brazilian Portuguese, Indonesian, and Korean options.

  • Google is rolling out 66 new regional voices in the Gemini app on Android and iOS, including U.S. English, U.K. English, India English, Hindi, Brazilian Portuguese, Indonesian, and Korean, with tailored options for each.
  • Google's description: each voice is tuned for natural cadences and everyday clarity rather than exaggerated accents, and is meant to capture familiar local rhythms, pitch, and inflection.
  • The selected voice carries across the app, including Gemini Live, language practice, and the Listen icon in chat for hearing a response read aloud.
  • Users can preview and choose a voice from the three-dot menu in Gemini Live, or from Gemini's voice settings in the app.
  • The rollout follows plans Google shared at Google I/O to bring local dialects to Gemini.
Sources:Google blog✓ official

Google — Gemini Notebook turns sources into interactive sites

Google

Gemini Notebook can turn notes and sources into interactive sites; the feature is rolling out globally on the web for users aged 18+.

  • Starting Nov 16, Gemini Notebook can turn a user's sources into interactive sites. Users type “make me a site” in the notebook chat and describe what they want to build.
  • Google's examples include visual explainers such as an interactive periodic table or a historical timeline, interactive dashboards for data such as quarterly sales performance, and project-management trackers for tasks and deadlines.
  • Once a site is built, users can open it in a new tab or share it with others.
  • The capability is rolling out globally on the web for Gemini Notebook users aged 18+, including Google Workspace subscribers, and is coming soon to the mobile app.
Sources:Google blog✓ official

Anthropic — Claude Code desktop adds iPhone Duo in the iOS simulator

Anthropic

Claude Code on desktop now supports iPhone Duo in the iOS simulator, in public beta, and can fold and unfold the device to test layouts.

  • Claude Code on desktop now supports iPhone Duo in the iOS simulator, in public beta, according to Anthropic's ClaudeDevs account.
  • Users can run their app on Duo in the simulator next to the conversation in the desktop app.
  • Claude can fold and unfold the device to test each layout.
Sources:ClaudeDevs on X✓ official

Liquid AI — d1.2 and d1.2-ultra omnimodal decision models

Liquid AI

Liquid AI released d1.2 and d1.2-ultra, hosted decision models that take text, images, and audio and return calibrated typed answers.

  • d1.2 and d1.2-ultra are hosted on the Liquid API, take text, images, and audio, and return calibrated typed answers with no generated tokens. They are called as d1.2 and d1.2-ultra with an API key from console.liquid.ai.
  • Pricing is $0.02 per million input tokens for d1.2 and $0.042 for d1.2-ultra, billed on input only. Liquid's documentation gives each model a 65,536-token input limit per question. An image is billed at 1.5 tokens per 32×32-pixel patch at the text rate, so a 1024×1024 image costs 1,536 tokens. Liquid says the models are coming to OpenRouter and Vercel soon, and that d1.1, d1-ultra, and the earlier d1 are deprecated.
  • The models add five tasks: reading a value (text, number, or date) with the characters it was taken from, or “not found”; finding every passage of a kind in a text; locating objects in a picture with one box or boxes for every instance; estimating a number with a range; and picking every option that applies, each with its own probability. Audio can be what the question is about, or the question itself can be spoken instead of typed.
  • Liquid's results: d1.2-ultra scores 75 on the Artificial Analysis Decision Index v2.1, the highest score and two points ahead of the previous best decision model; Liquid says no model scores higher than d1.2 for less money, and no model scores higher than d1.2-ultra at any price.
  • On calibration, across 13 benchmarks and 22,505 questions answered by all five compared models, Liquid reports a calibration error of 0.8 points for d1.2-ultra and 1.2 for d1.2, compared with 2.7 for Jev 1.14, 3.7 for Perplexity's Decider v1.3, and 7.2 for GPT-6 Luna on its Decisions API.
  • On Liquid's internal tests of the new tasks, d1.2-ultra scored 98% on reading a value from a receipt (CORD, SROIE), 94%/93% precision/recall on finding personal data (Nemotron PII), and 94% on locating a described object in a photo (RefCOCO family). On held-out recordings, both models answered 95% to 97% of questions about which sound they hear (ESC-50) or which language is spoken (FLEURS, 43 languages).
Sources:Liquid AI blog✓ officialLiquid AI docs✓ officialLiquid AI on X✓ official

Tencent — Hy4 flash released

Tencent

Tencent released Hy4 flash, a 224B-parameter multimodal model with 12B active parameters, a 1M-token context, and lower API prices than Hy4.

  • Hy4 flash has 224 billion total parameters, including 64 billion Engram parameters, with 12 billion active parameters. It supports multimodal inputs and a context length of up to 1 million tokens.
  • API list prices are about $0.070 per million input tokens, $0.281 per million output tokens, and $0.003 per million cached input tokens.
  • Tencent's description of the model: hybrid attention, LatentMoE, and Engram in pre-training; a KV cache 1/24 the size of Hy3 and 1/6 the size of Hy4; reinforcement learning that cuts average end-to-end task output tokens by 50%; and decoding speeds exceeding 150 tokens per second with megakernel, 4-bit quantization, and DFly optimizations.
  • Tencent's results: on its internal WorkIndex benchmark of typical WorkBuddy tasks, Hy4 flash's per-task cost was 1/13 that of Hy4 Preview, 1/20 that of Hy4, and 1/8 that of DeepSeek-V4.1-Flash-1017 at Max-tier peak-hour pricing. Tencent says Hy4 flash slightly outperforms the larger Hy4 Preview (770 billion total parameters, 49 billion active) in overall effectiveness.
  • In a blind evaluation with 85 internal Tencent WorkBuddy users across 259 tasks in 27 categories, Tencent says Hy4 flash reached user satisfaction close to Claude Haiku 5.5, with the lowest per-task cost and the shortest task completion time among the models evaluated.
  • Hy4 flash is available on Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima, and Marvis. Access is free on WorkBuddy through November 30, with 50% off standard pricing on TokenHub. Tencent later posted that the model is also live on OpenRouter at a limited-time 50% discount ($0.035 input, $0.1405 output, and $0.0015 cache read per million tokens). Tencent plans to open-source Hy4 flash and recommends it for everyday tasks, with Hy4 for long-horizon software engineering.
Sources:Tencent✓ officialTencent Hy✓ officialTencent Hunyuan on X✓ officialTencent Hunyuan on X (OpenRouter)✓ official

Cohere — Video Intelligence in Compass enters private preview

Cohere

Cohere added Video Intelligence to Compass, so agents can retrieve timestamped video moments alongside documents, now in private preview.

  • Video Intelligence in Compass is in private preview. Cohere is recruiting design partners, who can deploy it in Cohere's managed cloud or in their own environment.
  • Compass searches both what is said and what is shown on screen. Instead of returning a whole recording, it determines start and end times at search time and returns a timestamped clip with a transcript and a visual description, linked back to the original footage.
  • Video is retrieved from the same permission-aware index as documents, so applications and agents can pull footage alongside related pages, slides, or sheets in search, RAG, and agent workflows. An MCP server lets agents reason over on-screen material such as slides, charts, and demonstrations.
  • Cohere's results: it evaluated Compass against seven video embedding models across 11 datasets (six public benchmarks and five internal enterprise video datasets) and 11,133 queries. Compass ranked first on average, with an nDCG@10 4.4 points ahead of the strongest alternative.
  • Cohere's reported leads by domain: 2.4 points ahead of Amazon Nova Multimodal Embeddings, averaged across six public moment and video retrieval benchmarks; 5.1 points ahead of Gemini Embedding 2 on soccer broadcast footage; and 2.0 points ahead of Qwen3-VL-Embedding-8B on business meetings, lectures, and workplace-safety training.
Sources:Cohere blog✓ officialCohere on X✓ official

Subscribe for updates

Subscribe to get AI 24 Updates by email. You'll receive a confirmation email first.

We use your email address only to send the AI 24 Updates newsletter. You will get a confirmation email first, and every newsletter has an unsubscribe link. Subscriptions are managed by Buttondown (USA). See our Privacy Policy.