ON AIR · On air
AI model and tool releases, explained in plain words: what changed, whether to switch, and who it is for.
OpenAI mid-tier model speeds up computer use, halves API price
🐰 Rabbit takeSuccessor to GPT-5.6 Terra, built for screen-reading desktop automation like filling forms and logging invoices, with roughly half the error rate and more than double the coding score. Standard API pricing drops to $1 input / $5 output per million tokens—50% lower than before—making it suited to high-volume automation. Available now to ChatGPT Plus/Pro/Enterprise/Edu users in Work and Codex.
Qwen integrates into Apple Intelligence for China
🐰 Rabbit takeAlibaba and Apple announced Qwen will power Apple Intelligence in China, plugging into Siri, Writing Tools and Visual Intelligence for text/image understanding and generation. Developer beta starts Nov 4 and the public rollout lands with iOS 27.3 in December, so it isn't live for consumers yet. Users opt in, and requests or replies won't be used for training.
Claude renames top tier to Ultra and drops the 5-hour limit
🐰 Rabbit takeAnthropic renamed Claude's $100 Max 5x to Max and $200 Max 20x to Ultra, with no price change. Verified Ultra accounts no longer hit the 5-hour rolling usage limit, so long coding or research sessions won't be cut off mid-task—though the weekly cap still applies. Aimed at heavy users.
OpenAI doubles Work and Codex limits, cancels Pro 200 cut
🐰 Rabbit takeAfter subscriber backlash, OpenAI doubled Work and Codex usage limits across Pro 100/200/500 and Business Premium, and scrapped the planned Pro 200 cut. A real win for heavy Codex users, though Plus limits are unchanged. Doubled limits don't mean twice the completed tasks.
Claude adds CarPlay voice, 24 new languages and multi-account switching
🐰 Rabbit takeClaude's mobile apps now support CarPlay voice mode on iPhone, add 24 more languages (33 total), multi-account switching, and side-by-side generated content on iPad. Voice is also faster and more accurate in noisy environments.
Create, preview and edit docs, PDFs and sheets right in Gemini chat
🐰 Rabbit takeFiles you generate in Gemini—docs, PDFs, sheets, slide decks and code—can now be edited right inside the chat, either full screen or by highlighting text and asking Gemini to revise it. It's free for all Gemini users; PDFs work on mobile and web, while other formats are web-only for now.
Doubao Work opens third-party models to enterprise customers
🐰 Rabbit takeDoubao Work now lets Team and Enterprise subscribers run third-party models including Kimi 3.1 Pro, GLM 5.3 and DeepSeek V4.1 Flash—on by default and billed against existing AI credits, with no separate vendor procurement.
GitHub adds automatic multi-model routing for coding
🐰 Rabbit takeInstead of picking a model per task, Fusion auto-selects and orchestrates one or more models across Standard/High/Max effort levels. GitHub says Fusion Max scores within two points of Claude Opus 5.5 Max on the Coding Agent Index at about half the cost and 40% less time. Available in VS Code, the Copilot app, CLI and JetBrains; paid users get 30% off model costs through November.
ElevenLabs upgrades to SFX v3 sound effects model
🐰 Rabbit takeElevenLabs' SFX v3 sound effects model follows multi-sound prompts more reliably, improves music and vocals, and produces cleaner audio than SFX v2, available now in ElevenCreative and the API.
Suno launches official MCP to make music in AI chats
🐰 Rabbit takeSuno's first official MCP server (early access) generates, remixes and covers songs directly inside Claude, ChatGPT, Gemini, Cursor and Claude Code, using your existing Suno credits and saving to your library. It's gated behind a signup for now.
ChatGPT adds iCloud Mail and Calendar plugins
🐰 Rabbit takeChatGPT can now connect to iCloud Mail and Calendar to search, draft and send emails and read or update events, on all plans (including free) across web, desktop and iOS, using an Apple app-specific password. Useful for Apple ecosystem users, with the option to revoke access and opt out of model training.
Qwen app adds question search, photo answers and a study log
🐰 Rabbit takeQwen's app adds precise question search with free printing, step-by-step photo answers with an AI tutor, and auto-archived mistakes with weak-point analysis. Useful for K-12 students and parents; core features are free.
Runway lets you build custom creative apps from a description
🐰 Rabbit takeApp Studio turns a plain-language description into a shareable mini-app built on Runway's image, video and audio models plus your Brand Kits and workflows. Best for teams with repeatable creative pipelines (ads, film, e-commerce); included in paid plans and draws on existing credits.
Jimeng adds a one-stop short drama creation studio
🐰 Rabbit takeJimeng's Short Drama Studio bundles script writing, character/scene assets, storyboards, video generation and post-editing into one place, with 100k-character script parsing, up to 100-episode generation and team tools. Aimed at short-drama, adaptation and localization creators; some features offer free trials.
A 3-trillion-parameter open model with 1M context
🐰 Rabbit takeA new high-water mark for open models: stronger at long agent tasks, coding, spreadsheets and finance work, and it can turn text or sketches into 3D models. Works in China without a VPN. Full weights are due by November 4.
Cursor agents can now click and type across any desktop app
🐰 Rabbit takeCursor's coding agent can now click, type, and capture screens in any desktop app—not just the browser—so it can open the iOS Simulator, test the feature it just built, and revise the code in the same task. Most useful for mobile and GUI-app developers, it closes the build-test loop. It ships as a desktop-app update with no pricing change; on macOS you grant Accessibility and Screen Recording to the helper app.
Canva launches its first in-house image generation model
🐰 Rabbit takeCanva's first in-house image model is tuned for posters, menus and social graphics, with strong layout and typography — 2K output, 30+ aspect ratios and text in 55 languages. It's live in Dream Lab for all users (free tier has usage limits) and via a public API.
Anthropic launches Claude Duet, a proactive personal AI agent
🐰 Rabbit takeAnthropic's new personal agent reviews connected apps, surfaces what needs attention, and keeps working in the background on scheduling, forecasts and web apps — asking permission before acting on your behalf. It runs on Claude Opus 5.5 and is in a waitlisted research preview for US Pro/Max subscribers at no extra cost.
Kling 4.0 API launches with up to 30-second generation
🐰 Rabbit takeKuaishou's Kling AI has opened its new Kling 4.0 video model to developers worldwide via API. It brings steadier large-motion and complex camera moves, up to 30 seconds of native generation, up to 10 keyframes, and 4K 10-bit HDR output, with integrations across Magnific, Meitu and other platforms.
Shorter, faster answers; web and app updated too
🐰 Rabbit takeA polish of V4.1 Flash rather than a new model: tighter answers, faster task completion and better benchmarks. Prices are unchanged and the web and app switched over automatically. DeepSeek also opened its own search API.
ChatGPT Voice gets new voices and more natural accents
🐰 Rabbit takeChatGPT Voice moves to a new model with extra voices and more human-sounding speech. Included in existing plans at no extra cost; most noticeable if you practise speaking or use voice on the go.
Gemini 4 arrives in the app: more direct, much better at code
🐰 Rabbit takeGoogle's new flagship: less hedging and filler, clearer progress on long tasks, much stronger coding, plus a Tasks mode that keeps running while you are away and links to 40+ apps. Google AI Pro and Ultra subscribers only for now.
The fastest small model in the Claude family
🐰 Rabbit takeBuilt for speed and low cost: summaries, bulk processing and browsing. Casual chat will not feel very different; developers running high-volume API jobs or Claude Code subagents gain the most.
Over 30% faster and up to 30% cheaper
🐰 Rabbit takeSonnet is the Claude most people use day to day. This release is noticeably faster and cheaper, and on some tests it even beats the pricier Opus 5.5. For everyday writing and coding it is the value pick.
More expressive voices in 90+ languages
🐰 Rabbit takeA new generation of text-to-speech with richer emotion and tone, plus a low-latency Turbo version for voice agents. Worth another try for audiobooks and video voice-overs.
Stronger at coding and agents, 40% cheaper to run
🐰 Rabbit takeThe first Claude 5.5 model leads on agentic coding and knowledge work, and costs about 40% less to run than Opus 5 on typical workloads. Heavy Claude Code users benefit most.
Two cheaper GPT-6 tiers at half the API price
🐰 Rabbit takeSol targets coding and agents, Luna high-volume summarising and extraction. API prices are roughly half the previous generation, with fewer factual mistakes. Live in paid ChatGPT plans and Codex; Luna also reaches free desktop users.
A bigger base model at the same price
🐰 Rabbit takeA larger base model and a longer RL run than 4.6, with 500K context. Same price, and Grok users get it automatically.
Anthropic's most capable model for coding and knowledge work
🐰 Rabbit takeThe most capable model in the Claude family, aimed at the hardest coding and research work, and priced accordingly. A gated Mythos version is offered to vetted organisations. Most people should stay on Sonnet and switch up only for the toughest problems.
Native image understanding and lower API prices
🐰 Rabbit takeThe first Flash model with native multimodal input, so it can read images directly, plus lower API prices. V4 Flash is retired. Still free, fast and available in China without a VPN.
OpenAI's new flagship for computer use, browsing and software
🐰 Rabbit takeThe GPT-6 flagship, built to finish complex work end to end: operating a computer, browsing and building software, with over 1M tokens of context. Its strong cyber capabilities meant a limited rollout to vetted organisations before wider ChatGPT access.
The smartest Flash yet, same price
🐰 Rabbit takeGoogle calls it the smartest Flash model so far, with stronger coding and workflow skills, 1M context and the same introductory price as 3.7 Flash. It may use more reasoning tokens than its predecessor.
