← Reports

AI Radar

Saturday, May 16, 2026 · your own little newsletter of the cool AI stuff that's out there — fun things to play with and the developments worth knowing · AI Daily Brief (NLW) + Matt Wolfe energy, in your voice · enterprise/paper noise filtered, not the toys-only thing either

The all-encompassing read again, not the two-toys list. Full spread this week: new creative tools you can actually play with, the model/app news that matters, the speech & on-device stuff that's your long arc, the agent/coding-tool moves, and what's landing in the next few days.

Top picks

Anthropic puts Claude agents on a meter — June 15 changes how you run Claude Code

23/25

Anthropic pricing / policy affects your workflow · announced May 14, effective Jun 15 · axios.com · infoworld.com

The big one for how you work. From June 15, programmatic Claude use on subscription plans (Claude Code, Agent SDK, headless claude -p jobs) moves into a separate fixed monthly credit pool billed at API rates — no longer subsidized by the general subscription bucket, no rollover. NLW's read on the AI Daily Brief: the cheap-token era that made endless agent experimentation possible is ending; the developer reaction has been loud ("gaslighting," people threatening to jump to Codex).

Directly relevant: your stack runs a lot of headless/parallel Claude (the radar generator itself, briefs, the agent fleet). This isn't "play with it" — it's a budget change to model before June 15 so the automated pipelines don't quietly blow a credit pool. Front-load heavy agent work before the cutover; watch for the metered-usage dashboard when it lands.

jun 15 cutover agent sdk metered model before it hits

Wispr Flow — the brain-dump-to-clean-text app, on all four platforms now

22/25

speech / dictation play with it Bidet-shaped · trending on Product Hunt this week · wisprflow.ai

Basically a commercial cut of the Bidet thesis, so worth a hands-on look. You speak, it transcribes, then layers strip "um/uh/like," fix punctuation, repair your backtracking, and adapt the writing style to the app you're in (casual in Slack, structured email in Gmail) — the LLM-error-correction-after-ASR pattern, generalized past your v18.9 regex. Now on Mac/Windows/iOS/Android, 100+ languages, even works whispering. The catch: it's cloud-only, no offline (Bidet's on-device angle is still the differentiator).

Try it for a day as a benchmark for what "good" feels like, and as a reference for Bidet's cleaning model — the per-app style adaptation is a feature worth stealing for a future Bidet output mode. Paid tier ~$15/mo (optional; free tier to test the feel).

asr + llm cleanup per-app style bidet reference

Google I/O is Tuesday — new Gemini, Android XR glasses, agentic Android

21/25

Google what's coming glasses / on-device · keynote May 19, 1 PM ET · androidauthority.com

The big "what's-next" of the week. Google I/O 2026 runs May 19–20; the keynote is Tuesday 1 PM ET (io.google streams free). Expected: a major Gemini model bump, a real preview of Android XR smart glasses (live translation, heads-up notifications, Gemini built in), and "Gemini Intelligence" — an agentic AI push baked into Android.

Two threads for you: a new Gemini tier matters because Bidet routes to Gemini for cleaning, and Android XR glasses are the obvious "Computer voice trigger → glasses speak" successor to the Ray-Bans pattern — worth watching the keynote with that lens. Next radar carries the actual announcements; this is the heads-up.

tuesday keynote gemini bump xr glasses

NVIDIA Nemotron 3 Nano Omni — one open model that hears, sees and reasons

20/25

open weights speech / audio on-device-ish · released late Apr, traction now · blogs.nvidia.com · huggingface.co

A 30B mixture-of-experts model that activates only ~3B params per pass and folds vision + audio encoders into one stack — no separate STT model. Tops six leaderboards on document, audio and video understanding with ~9× the throughput of other open omni models. Open weights on Hugging Face and OpenRouter.

Same "one model, audio + text, no separate Whisper stage" direction as the Gemma 4 E2B/E4B bet, but a credible open alternative — relevant if you ever want a non-Gemma fallback for the Bidet speech path. 3B active means it's plausibly edge-capable; a Gemma comparison, not a swap today.

native audio moe 30b/3b gemma alternative

Also on the radar

ElevenLabs ElevenMusic — free AI music app, 7 songs/day from a prompt

18/25

audio / music free, play with it · officechai.com · techcrunch.com

ElevenLabs entered the Suno/Udio music war with ElevenMusic (iOS) — type a natural-language prompt, get a track; currently free, ~7 songs/day. Built on licensed catalog (Merlin, Kobalt), and as you'd expect the vocals are the most natural of any generator right now.

Pure play-with-it: intro music for a Bidet/contest video, or just messing around on the patio. Free tier, no reason not to poke it. (Suno still leads on speed/features, Udio on fidelity — ElevenLabs wins on voice.)

prompt to song free video music

Claude Opus 4.7 — hand off your hardest coding without babysitting it

18/25

Anthropic model launch your daily driver · anthropic.com

Opus 4.7 is GA and the gains are concentrated where you feel them: the hardest software-engineering tasks — the kind that used to need close supervision — can now be handed off with confidence. Notable step over 4.6 on difficult, multi-step work.

This is the model running your agents and the Bidet/contest builds. Practical angle given the pricing card above: better one-shot success means fewer retry loops — which matters more once headless usage is metered June 15. Lean on it for the gnarly Bidet work now.

hardest tasks fewer retries

The 2026 open STT board — Canary-Qwen 2.5B on top, Moonshine for tiny

17/25

speech / STT on-device Bidet pipeline · gladia.io · northflank.com

Current open-ASR landscape, your long arc: NVIDIA Canary-Qwen 2.5B leads the HF Open ASR leaderboard at 5.63% WER, IBM Granite Speech 3.3 8B close behind at 5.85%, and Moonshine still owns the resource-constrained niche (runs on Pi-class / mobile). None ship diarization — pair with pyannote/WhisperX or the new SpeakerKit (Core ML, Pyannote v4) for who-said-what.

Moonshine stays the bet for STT on lower-end phones than the Pixel 8 Pro; diarization is still a separate stage to bolt on for the lecture-recording use case. Nothing to swap today, but this is the board to track for the next Bidet on-device upgrade.

canary-qwen moonshine diarization separate

Luma Ray3 — "master the best shots" into 4K HDR, contest-relevant

17/25

video play with it Bidet contest · lumalabs.ai · pixflow.net

Luma's Ray3 builds on Ray2 with stronger realism, physics and character consistency, plus Hi-Fi Diffusion that "masters" your best takes into production-ready 4K HDR. The video field is hot right now (Runway Gen-4.5 tops the text-to-video Elo board; Kling 3.0 added native 4K + storyboard + lip-sync).

Directly your contest lane — Luma is the platform you're already pointed at for the Bidet pitch video. The "master a take to high-fidelity" step is the useful new bit for one clean hero shot. Runway/Kling are the comparison bench if Luma isn't landing the look.

4k hdr character consistency contest tool

Coding-agent shuffle — OpenAI Codex mobile approvals, Copilot goes credits

15/25

agent / coding tools worth knowing · sdtimes.com · codenewsletter.ai

The coding-agent field moved this week: OpenAI added mobile supervision for Codex (review & approve agent work from the ChatGPT phone app, May 14), GitHub Copilot is shifting to AI-credits billing June 1, and ServiceNow's Build Agent now plugs into Cursor/Windsurf/Claude Code/Copilot. The field is converging on metered, governed, surface-agnostic agents (same theme as the Anthropic pricing card).

You're a Claude Code person so no action — but "approve from your phone" is the same shape as remote-controlling your agents from the Pixel, and the industry-wide move to credit metering confirms the June 15 Anthropic change isn't a one-off. Context, not a switch.

codex mobile copilot credits metering trend

NotebookLM Cinematic Video Overviews — notes → narrated mini-movie

14/25

creative tool teaching fit · blog.google · xda-developers.com

Not brand-new (rolled out March) but now on all Google AI Pro/Ultra accounts, so it's actually reachable: upload sources, click generate, get a 1–3 min narrated explainer video (Gemini + Imagen + Veo) — no script, no editing, up to 20/day.

Teaching angle: turn a unit's source docs into a short explainer for a class, or a Legacy Soil / Breezy Farms one-pager into a quick walkthrough video. Lowest-effort "notes become video" path right now — worth one test run to see if the output is classroom-usable. (Needs the AI Pro plan you may already have.)

notes to video teaching zero edit

Product Hunt this week — ElevenCreative Flows, Stitch 2.0, self-modifying apps

13/25

new launches play / browse · producthunt.com

The week's creator-AI churn worth a scroll: ElevenCreative Flows (chained image/video/audio pipelines) and Templates, Stitch 2.0 (prompt → editable UI), Arcana Labs (image+video), plus the weirder end — "Stella," a self-modifying desktop app.

The "go play" bucket. ElevenCreative Flows is the one to actually open — chaining audio+video steps is the kind of one-pass content pipeline that could speed up Bidet/contest media without you stitching tools by hand. The rest is good Saturday-tinkering material.

elevencreative flows tinker

Gemini 3.1 Flash TTS — voice with embedded tone/pacing tags

13/25

voice / TTS speech-adjacent · blog.google · llm-stats.com

Google's preview TTS model (Gemini 3.1 Flash TTS) lets you embed natural-language audio tags inline to steer tone, pacing, accent and expression — the opposite end of your pipe (text→speech rather than speech→text), and a cheaper Gemini-side narration option than the premium voice tools.

On the radar because Bidet/contest narration and the /brief → Ray-Bans TTS path both want controllable voice. Worth a quick test as a free-tier-friendly narration option. Watch for an updated version dropping at I/O Tuesday.

inline audio tags narration

Coming up on the radar

Cut with reason

Pipeline status

This run