MODEL6
A new foundation model — the brain.
Shifts what's possible. New ceiling on reasoning, coding, context.
latestDeepSeek V4.1 FlashDeepSeek
← stanwood.dev
AI RADAR
8 trackedWhat just shipped in AI that's actually worth knowing about.
8 launches·8 orgs·Sep 1 – Sep 10·6 models·2 tools
The Pulse
last 7 days · last 30 days4
this week
vs 4 prior
0d
since last drop
shipped today
DeepSeek
on a tear
1 in 30d
MODEL
most this month
6 of 8
The Field
what we trackPRODUCT0
A consumer app or feature.
How non-builders actually feel AI in their day.
watchingnone in feed yet — when one ships, it lands here
TOOL2
A developer API, SDK, or capability.
What you can build with this week — the new primitives.
latestDesert Ant LabsDesert Ant Labs
INFRA0
Hardware, platforms, the backbone.
Who runs what at what price — the rails everyone shares.
watchingnone in feed yet — when one ships, it lands here
The Labs
nine you'll see mostMost launches in this feed come from the same handful of labs. Here's who they are, where they're based, and what they're actually known for — so the org chips below stop being acronym soup.
OpenAI
Set the pace for the modern wave — ChatGPT made conversational AI a household tool overnight.
closed weights1 in feed
GPT-5 · ChatGPT · Sora · Realtime API
Anthropic
Founded by ex-OpenAI safety researchers; Claude is the model coding tools quietly standardize on.
closed weights1 in feed
Claude Sonnet · Claude Opus · MCP · Skills
Google
DeepMind + Brain merged into the Gemini team; the only lab with the chips, the data, and the products.
mixed1 in feed
Gemini 3 · AI Studio · Vertex · Veo
Meta
Open-weights leader in the West — Llama set the floor every other open lab is measured against.
open weights
Llama · Movie Gen · Ray-Ban Display
xAI
Musk's lab. Massive Memphis training cluster, edgier brand voice, integrated tightly with X.
mixed
Grok · Colossus supercluster
Mistral
Europe's flagship lab — punches above its weight on small, efficient, mostly open-weights models.
open weights
Mistral Large · Mixtral · Codestral
DeepSeek
Sparked the 2025 cost-collapse — frontier-class reasoning at a fraction of the price, weights and all.
open weights1 in feed
DeepSeek-R1 · V3 · open weights
Alibaba
Qwen team ships open-weights models at every size with ruthless cadence — the workhorse of Chinese AI.
open weights
Qwen 3 · Qwen-VL · Qwen-Coder
NVIDIA
Sells the picks and shovels. Every other lab on this list trains on its chips — H100, B200, GB200, and counting.
infrastructure
Blackwell · CUDA · NeMo · Nemotron
Plenty of others ship too — Cursor, GitHub, Cloudflare, Midjourney, Hugging Face, Runway, ByteDance, MiniMax, Z.ai. They show up in the chips below as their launches land.
Signal vs. Noise
six tells in AI launches · spot the spin fastMost launches are real. Some are framing. These are the six tells that separate "actually ships" from "press release with a roadmap" — handy when a headline beats the feed below to your inbox.
Benchmark cherry-picking
watch forCharts that only show the benchmarks they win. Comparisons against last year's model, not this week's.
what it meansReal on a narrow slice — often weaker on the harder evals (GPQA, ARC-AGI, Humanity's Last Exam) that didn't make the slide.
Available soon vs. now
watch for"Rolling out over the coming weeks." Waitlists. A blog post but no API page, no pricing, no model ID.
what it meansMarketing landed today; the product might land later. Treat the date as when it became real to journalists, not to you.
Cheaper, dressed as a leap
watch forHeadlines built on "10× cheaper" or "3× faster" with no matching jump on capability evals.
what it meansBig win if you're spending on inference. Not a new ceiling — same model territory at a better price.
Demo magic vs. ship
watch forHighlight reels, narrator voiceover, "hand-picked examples," only logged-out demos.
what it meansStage version ≠ API version. The cherry-picked scene is the ceiling, not the median run you'll get.
Open-weights ≠ open-source
watch for"Llama community license," commercial caps ("700M monthly users"), custom acceptable-use clauses.
what it meansFree to download. Not free to ship. Read the license before betting a product on a model.
Model card vs. blog post
watch forPerformance claims with no methodology. "Internal evals" footnotes. No system card on launch day.
what it meansIf they didn't publish the eval setup, the numbers are vibes. Wait a week for independent runs (Artificial Analysis, LMSYS).
None of these mean a launch is fake — just that the spin is doing work. The feed below tries to flag the spin in the summary when it spots it.
MODELtoday
DeepSeek V4.1 Flash
DeepSeek — Run a 1M-context multimodal model for pennies — $0.30 per million input tokens at peak, half that off-peak. It beats the older V4 Pro on speed and cost, and the 485B weights are up on Hugging Face.
read more →8SEP
TUEChatGPT Images 2.5MODEL
OpenAI — Generate and edit images with much better likeness for people and pets, up to 50% faster. Rolling out to every ChatGPT tier plus Codex, with a @Sketch tool for drawing references and two new API models, Flare and Sunburst.
Desert Ant LabsTOOL
Desert Ant Labs — Ship transcription, audio cleanup, PII redaction, and language ID that run entirely on-device, via Swift, Kotlin, or JavaScript SDKs. 18 models, no token billing, free up to 100,000 monthly active devices.
3SEP
THUK2 HorizonMODEL
Institute of Foundation Models — Run a capable open model locally at 0.9B, 3.7B, or 7B, or self-host up to 375B. Apache 2.0 with weights, training data, and code published, plus day-zero Ollama, vLLM, and GGUF support.
Hermes DesktopTOOL
Nous Research — Set up a local model in one click — the app reads your RAM, VRAM, and GPU, picks a model that fits, and starts a llama.cpp server. MIT-licensed and free on Mac, Windows, and Linux.
About this radar
A hand-curated log of AI launches that actually move the field — new flagship models, agentic products, developer APIs, and infrastructure shifts. No press releases. No minor patches. No prerelease beta noise.