AI Tools This Week: Glaze Mac Apps, GPT-Live Voice, Stanley Studio, and Grok 4.5

The most interesting AI launches this week are not asking you to learn another interface. They are trying to remove one.

Glaze turns a description into a Mac app. GPT-Live tries to remove the rigid turn-taking from voice assistants. Willow replaces typing. Stanley Studio replaces part of the video timeline. Claude Cowork mobile turns the phone into an approval surface. Grok 4.5 and GPT-5.6 make the model layer cheaper and more competitive underneath all of it.

Video credit: The Next New Thing, with Andrew Warner and Corey Ganim. Follow the channel on YouTube and Andrew on LinkedIn.

Source Note

This post uses Andrew Warner and Corey Ganim's roundup as the discovery layer, then checks product claims against official Glaze, OpenAI, ElevenLabs, Willow, Stanley Studio, Anthropic, SpaceXAI, and Claude documentation. X posts are labeled as creator, founder, or user commentary rather than official product guarantees.

The supplied X links are included exactly, but one needs a correction: Elon Musk's second Grok post discusses a possible 1M-context upgrade, not a benchmark result. Treat "probably by next week" as a forecast until the official model docs change.

Link Map

The Pattern: AI Is Removing Interfaces

These launches look unrelated until you ask what friction each one removes.

• Glaze: less Xcode, packaging, and desktop-app boilerplate.
• GPT-Live: less waiting for clean conversational turns.
• Willow: less typing and cleanup.
• Stanley Studio: less timeline dragging and manual caption work.
• Cowork mobile: less dependence on sitting at the desktop while an agent works.
• Claude video demos: less repetitive screen recording and editing.

The products that win will not merely have a smart model. They will translate intent into a finished artifact while preserving enough control for a human to correct the result.

Glaze: Personal Mac Apps Without Starting In Xcode

Glaze by Raycast lets users describe a desktop application and build it through conversation. The resulting apps run locally, can work offline, integrate with operating-system features, and can connect to local files, hardware, APIs, and AI services.

That makes Glaze more interesting than another browser app builder. A desktop app can use keyboard shortcuts, menu-bar controls, background processes, file-system access, and local data. Glaze says users own the app, code, and content they create, and can publish to a public store or share privately through paid plans.

The current limit matters: Glaze requires Apple Silicon and macOS Tahoe. Windows and Linux are planned but not currently supported. I would test it first on one narrow utility:

• a menu-bar client timer,
• a file-renaming utility,
• a local image or PDF organizer,
• a project status dashboard,
• a private wrapper around one internal API.

Do not start with a mission-critical company app. First inspect the generated code, permission model, update path, portability, and how the app behaves when Glaze is unavailable.

GPT-Live And Willow: Voice Becomes An Interface

Official OpenAI demo. Read the companion GPT-Live release or the JQ AI SYSTEMS GPT-Live builder guide.

OpenAI says GPT-Live uses a full-duplex architecture: it continuously processes input while producing output, so it can listen, speak, pause, interrupt, or invoke a tool without waiting for rigid turns. For search, reasoning, or more complex work, GPT-Live delegates to a frontier model while maintaining the conversation. At launch, that background model is GPT-5.5.

The Husk X clip is best treated as a user stress test, not as a formal "issue report." Voice systems should be tested with accents, interruptions, background audio, silence, emotional speech, translation, unsafe requests, and accidental microphone activation before anyone trusts them in a real workflow.

Willow Voice attacks a simpler problem: getting text into any application quickly. It supports Mac, Windows, and iOS, with personal terminology, formatting, shortcuts, transcript history, and style matching. Its free plan includes 2,000 words per week. The official help center lists Pro at $15 per month before annual discounts.

Willow is the easiest tool in this roundup to evaluate. Dictate the same ten emails, prompts, notes, and technical phrases with Willow and your current method. Count correction time, not just transcription speed.

ElevenLabs Music And Stanley Studio

ElevenLabs Music v2 combines prompt-based composition with section editing, inpainting, reference matching, multilingual vocals, and API access. ElevenLabs says the model is trained on licensed data and generated tracks have commercial-use paths, but the plan and medium matter. Its public page says self-serve commercial use excludes film, television, and studio games, while Enterprise terms cover broader use. Check the model-specific terms before publishing.

The more interesting workflow is not "make one song." It is programmatic music inside a content pipeline: generate several beds, match duration and mood, revise only the weak section, then route the final track into a video editor.

Stanley Studio provides that editing layer. Upload clips and describe the result. Its official examples show silence removal, best-take selection, captions, punch-ins, background music, text hooks, B-roll, and 9:16 reframing from one natural-language edit brief.

The promise is not that timelines disappear for professional editors. It is that creators can delegate a strong first assembly and spend their attention on pacing, story, brand, and the keeper moments. The review checklist should still cover safe zones, caption accuracy, music rights, abrupt cuts, visual continuity, and whether the edit says what the creator intended.

Claude Cowork Mobile: Approval Moves To The Phone

Video credit: Rob The AI Guy. Product facts are checked against Anthropic's official Cowork announcement.

Anthropic says Claude Cowork is rolling out to web and mobile over several weeks, starting with Max users. Remote sessions and files follow the user's Claude account, work can continue after the laptop closes, and scheduled tasks can run without a device online.

The phone is not replacing the full desktop experience. Anthropic's help page says local file access, local connectors, browser use, and computer use from web or mobile still work through Claude Desktop. The mobile value is continuity: start work, redirect it, inspect progress, review the artifact, and approve what happens next.

That is a healthier agent interface than pretending autonomy means zero supervision. Mobile should make the approval loop easier, not remove it.

Grok 4.5 And GPT-5.6: The Model Layer Gets Cheaper

Video credit: Theo - t3.gg. See the JQ AI SYSTEMS Grok 4.5 builder analysis for the deeper benchmark and pricing breakdown.

SpaceXAI prices Grok 4.5 at $2 per million input tokens and $6 per million output tokens. The official page positions it for coding, agentic tasks, knowledge work, Grok Build, Cursor, and API use. It also says the model is not yet available in the EU and expects availability in mid-July.

Elon Musk's performance assessment and context-window forecast are useful launch signals, but the official model page and your own evals should drive production decisions.

Dan Shipper's two GPT-5.6 posts offer a complementary operator view. In the writing post, he reports strong one-shot marketing copy. In the broader analysis, he describes GPT-5.6 as a strong balance of power, speed, and day-to-day usefulness while reserving Fable for work that needs a more unusual level of open-ended intelligence.

Those are thoughtful observations, not universal rankings. The useful routing rule remains: test cost per accepted result on your own coding, writing, and knowledge-work tasks.

Claude Code Can Generate The Demo Instead Of Recording It

Ole Lehmann's Claude video demo shows a small but clever workflow: instead of repeatedly recording a cursor typing prompts, ask Claude Code to generate the animation.

This is useful for product demos, course clips, landing-page loops, onboarding animations, and B-roll where the interaction is deterministic. A generated demo is easier to revise, resize, translate, and keep visually consistent than a manual screen recording.

The caution is honesty. Generated UI demos should not imply a feature, speed, or outcome the product cannot actually reproduce. Label conceptual animations when needed and keep real product verification separate.

What I Would Test First

The best launch is the one that removes a real bottleneck without creating a larger review, privacy, or reliability problem. This week, start with the interface you already dislike most.

Sources

• The Next New Thing: Free Mac apps, AI video editing, dictation, and cool AI tools launched this week
• The Next New Thing and Andrew Warner
• Glaze by Raycast
• OpenAI: The next generation of ChatGPT Voice
• OpenAI: Introducing GPT-Live
• Husk: ChatGPT Voice user stress test
• ElevenLabs Music v2
• Willow Voice and official pricing overview
• Stanley Studio
• Rob The AI Guy: Claude Cowork mobile access and upgrades
• Anthropic: Claude Cowork on web and mobile
• Claude Help Center: Cowork across web, desktop, and mobile
• SpaceXAI: Introducing Grok 4.5
• Elon Musk: Grok 4.5 internal assessment
• Elon Musk: possible Grok 4.5 1M-context upgrade
• Theo - t3.gg: Oh no, the new Grok model is good
• OpenAI: GPT-5.6
• Dan Shipper: GPT-5.6 writing review
• Dan Shipper: GPT-5.6 day-to-day analysis
• Ole Lehmann: Claude-generated video demos