Block's Speech-Only Voice Model, Anthropic's IPO Filing, and More
The release I keep coming back to today is Dessa, a new voice model from Block. It's trained only on speech, with no text at all. That shows in how it behaves: it's full-duplex, so it listens and talks at the same time, and it laughs mid-sentence and backchannels the way people do on a call. (Block)
It also runs faster than real time on an M4 Max MacBook Pro, which makes it a local model, not just a data-center demo. Block is giving it to safety researchers first and has opened a public waitlist for everyone else. A model this human-sounding is exactly the kind of thing I'm glad to see go to safety researchers before it goes wide. (TechCrunch)
What Else Happened
Anthropic's S-1 could go public as soon as Monday. Reports say the public IPO filing may land ahead of a possible November roadshow. Over the weekend, two people behind Claude Code also announced they've left: Thariq Shihipar and Anthony Morris, who created Claude Code on desktop. (X news summary, Thariq Shihipar, Anthony Morris)
Kimi K3.1 Pro has shipped. It's a 2.8T-parameter MoE with a 1M-token context at $2 per million input tokens. It ranks 4th on the Artificial Analysis Intelligence Index and #8 in Code Arena WebDev, and open weights are promised by November 4. The launch also drew backlash over an undisclosed creator campaign pushing it as "open-source Opus 5.5," which is a good reminder to read the benchmarks yourself. (Vercel AI Gateway, the backlash)
The Claude desktop app got a design refresh. New icons, a simplified type system, and better contrast, and it went from sketch to ship in under three weeks with Claude helping. It's still rolling out, and some Windows users already want the old dark theme back. (Meaghan Choi)
GPT-6.1 Terra is showing up in partner configs. Developers spotted the mid-tier model in Microsoft and GitHub Copilot test setups, slotted between Luna and Sol. OpenAI hasn't confirmed it, so treat this as a sighting, not a launch. (X news summary)
Launches and Demos
- Ray Fernando's deploy watcher – a Cursor Projects agent that ships his app, watches production for an hour, and rolls back the tags automatically if anything breaks.
- Pegpix – a free fuse-bead pattern maker built with Opus 5.5 by an art director who doesn't write code. Draw, turn a photo into beads, or have an agent design one, then print it at actual size.
- HyperFrames – HeyGen's video-storytelling tool, designed with Gemini 3 Flash. Their lesson: the thinnest wrapper, plain HTML plus a few data attributes, beat bigger prompts and more skills.
- Beamline UI over MCP – ask your coding agent for a client onboarding form and get back real components: steppers with units, tappable goal cards, a slider, and a date picker.
- Agent Canvas – a free, MIT-licensed tool that shows you the real running UI. Point at an element and copy a precise change request into Claude Code or Codex.
- RustGIS v0.4.0 – open-source map software that runs in the browser, now drivable by an agent with undo on every step. This release adds 3D terrain, Moon and Mars maps, and Esri geodatabases without GDAL.
Accounts Worth Following
- Danylo Pravda (@danylo_pravda) – a designer building a design system made for both you and your coding agent, sharing the builds as he goes.
- Matt (@matsugfx) – a design engineer focused on design systems who writes candidly about how AI is changing designers' work.
- James Russo (@Rames_Jusso) – works on HyperFrames at HeyGen and shares practical lessons about building agent-friendly video tools.
What I'm Reading
- Dessa – Block's page for its speech-only voice model, with the waitlist.
- Block's new voice model sounds eerily human – TechCrunch's write-up of Dessa.
- The Claude desktop refresh – Meaghan Choi on going from sketch to ship in under three weeks.
- Kimi K3.1 Pro on Vercel AI Gateway – the launch post with availability details.
That's it for today. 👋