# Dan's Radar — Week of September 28, 2026 ## Executive summary A heavier week than usual for platform moves, with OpenAI's Dev Day setting the pace. OpenAI shipped **Dots**, its own personal assistant, alongside GPT-6.1 Sol, a Decisions API aimed at Jev, and **Sign in with ChatGPT**, which lets third-party apps ride a user's subscription and inference; Google answered with Gemini 4 **Argon**, rolled out to cyber defenders first and now top of the Vals Index. Instinct started pushing **sponsored products** to users, an early answer to how free agents get paid. NVIDIA met the recent agent-breakout reports with an **Open Agent Safety Platform** that sandboxes agents and enforces policy in silicon. On the money side, Anthropic's IPO figures circulated, a Stripe chart showed US AI payment volume near vertical while the **EU** stays flat, and a16z data put paying US households at 2%. ## Signals (12) ### 1. OpenAI Dev Day in one post: Dots personal assistant, GPT-6.1 Sol, a Decisions API aimed at Jev, $500/month Ultrafast, and Sign in with ChatGPT — 9.0 *X · Theo (@theo) · Tue* Theo's summary of OpenAI Dev Day: Dots, a personal assistant in the Muse and GrokBot mold; GPT-6.1 Sol, cheaper and stronger than 6; a Decisions API with vision positioned against Jev; a $500/month plan with 300 tok/s Ultrafast; Sign in with ChatGPT for third-party apps; and extensions that can be full apps. > Dan's take: everything OAI announced today https://x.com/theo/status/2104995863689142546?s=20 --- ### 2. NVIDIA proposes an Open Agent Safety Platform: OpenShell sandboxes plus in-silicon monitoring on BlueField-4, with a partner logo wall — 7.5 *Web · NVIDIA Technical Blog · Mon* NVIDIA published a reference design for agent safety: open-source OpenShell runs each agent in a kernel-isolated sandbox with verifiable policy, and NVIDIA Sentry on BlueField-4 DPUs enforces it out of band on the node's only path to the model. It cites recent agent breakouts at frontier labs and lists ecosystem partners. > Dan's take: we should be in here https://developer.nvidia.com/blog/nvidia-open-agent-safety-platform-a-reference-for-continuous-in-silicon-agent-monitoring --- ### 3. Instinct starts pushing sponsored products to users based on purchase history and browsing — 7.5 *X · Andrew Yeung (@andruyeung) · Tue* Andrew Yeung reports Instinct has begun promoting products to him, apparently sponsored and matched to his purchase history and internet activity, and reads it as the free product's monetization model taking shape. > Dan's take: ads is a viable/expected avenue to monetize agents; users surprised. https://x.com/andruyeung/status/2105076419076174239?s=20 --- ### 4. Google ships Gemini 4 Argon to cyber defenders first at $2 in / $10 out; it takes #1 on the Vals Index — 7.0 *X · Logan Kilpatrick (@OfficialLoganK) · Wed* Logan Kilpatrick introduced Gemini 4 Argon, Google's new frontier model, rolling out to cyber defenders first and more widely soon, at introductory pricing of $2 per million input and $10 output. Vals AI put it at #1 on its index at 68.9%, Gemini's first time on top. > Dan's take: Argon made some rounds and challenged everyone's benchmarks https://x.com/OfficialLoganK/status/2105388054274080946?s=20 --- ### 5. Anthropic IPO prospectus figures circulate: $4.59B revenue up 1,088%, $41.97B GAAP loss, $7.3B compute, $20.3B cash — 6.5 *X · Shay Boloor (@StockSavvyShay) · Mon* A viral post lists figures from Anthropic's IPO prospectus, per Reuters: $4.59B revenue (+1,088% YoY), a $41.97B GAAP net loss, $7.3B in compute at 58% of opex, and $20.3B cash. Community notes clarify the S-1 was filed confidentially in June, not that day. > Dan's take: there is no more cash in the market, though https://x.com/StockSavvyShay/status/2104715381780951305 --- ### 6. How Supabase became the database agents recommend by default, and where even it gets zero AI citations — 6.0 *X · mal (@mal_shaik) · Sun* mal traces why agents recommend Supabase first: 15+ launch weeks, comparison pages for every rival, 99k GitHub stars, and a page for every how-do-I query. Yet ChatGPT gives it zero citations for its own core features like row-level security and realtime when the question describes the job, not the category. > Dan's take: used by 60% of YC latest batch. they come associated with Vercel and Cloudflare is the enemy. https://x.com/mal_shaik/status/2106821061585359127?s=20 --- ### 7. Supabase Select 2026: Turso acquisition for agent-scale databases, Supabase Compute, an MCP server for your app, local without Docker — 5.5 *X · Sam Rome (@samromehq) · Fri* Supabase Select 2026 announcements: acquiring Turso for agent-scale databases, Declarative Schemas 2.0 with generated migrations, Supabase Compute for long-running services beside Postgres, local Supabase without Docker, an MCP server per app, Multigres HA with OrioleDB, and open Postgres benchmarks. > Dan's take: Supabase compute sounds fire! https://x.com/samromehq/status/2106106277613224165?s=20 --- ### 8. a16z data: 2% of US households pay for AI, 91% pay for streaming; 'X is the bubble inside the bubble' — 5.5 *X · Danni Chen (@DanniChenGTM) · Sat* Quoting a16z's State of Markets II, Danni Chen notes about 2.7M of 134.8M US households pay for AI (2%), against 91% paying for streaming and 55% for cloud storage, and warns that a GTM plan built on what an SF group chat pays for targets the 2%. > Dan's take: either a bubble, or comfortable time to develop for the long tail? https://x.com/DanniChenGTM/status/2106609065040208124?s=20 --- ### 9. C5R built a research facility run entirely by AI in 12 weeks; one model controls equipment, researchers, and inventory — 5.0 *X · donald (@donaldjewkes) · Thu* C5R says it built a lab in 12 weeks where AI designs, runs, and observes experiments across biology, chemistry, and materials science, and launched SciUniverse, a benchmark for real-world science. A contributor describes two weeks inside it, where a single Astra-powered model can be prompted to run the whole facility. > Dan's take: just don't get caught up when the token price goes up https://x.com/donaldjewkes/status/2103214063832694819?s=20 --- ### 10. Anthropic: robots can already do 74% of physical tasks but are cost-competitive for 0.3% of work; drivers and packers first — 5.0 *Web · Anthropic Economics (Legate-Yang, Massenkoff) · Wed* Anthropic's economics team built a robot exposure index from 7,594 physical O*NET tasks: today's robots can do 74% of physical work (34% of all hours), but are cheaper than humans for only 0.3% of it, with 40 years to reach 10% at historical price declines. Robots plus LLMs expose about 80% of work. > Dan's take: most exposed age is 38 https://www.anthropic.com/research/what-work-can-robots-do --- ### 11. Chart: monthly payment volume of AI businesses on Stripe, US curve near vertical since 2025, EU flat since 2023 — 4.0 *Web · Marius Comper (Facebook) via eudata.vercel.app · Tue* A chart shared on Facebook with the caption 'Ouch' plots monthly Stripe payment volume of AI businesses by HQ country from January 2023 to 2026: the US line rises roughly tenfold with a steep climb through 2026, while the EU line stays almost flat along the bottom axis. > Dan's take: pacing is good but the EU will catch up https://eudata.vercel.app/ --- ### 12. 45+ SF tech events for the week of Sept 28 to Oct 5, filtered to five picks — 2.0 *X · Michelle Fang (@michelleefang) · Mon* Michelle Fang's weekly edit lists more than 45 in-person SF tech events for Sept 28 to Oct 5, narrowed to five picks with reasons, plus a form to submit events for next week's edition. > Dan's take: this is a solid list of events in SV https://x.com/michelleefang/status/2104678199443800391 ---