358 tweets liked between September 20 and September 29, filtered down to what’s worth knowing (here’s how the pipeline works).

AI for Everyone

OpenAI Launched Dots at DevDay, and My Feed Shrugged

OpenAI introduced dots, always-on agents powered by GPT-6 Astra, its top model. Each dot has its own cloud computer and browser, works across 4,000+ apps through ChatGPT plugins, and runs around the clock, and you set what it does alone, when it asks first, and what it never does. OpenAI lists Pro, Business Premium, and Enterprise as the plans, but as of 5:45 pm Eastern on Sep 29 nobody I could find had access yet. OpenAI’s Tibo Sottiaux says it won’t draw down your usage limits. Sam Altman had teased “a new thing” the night before, and the reactions I liked afterward were mostly flat. Ibrahim Ahmed called DevDay “super underwhelming” and short on ambition. NIK summed up dots as costing $200 to do what Muse, Meta’s agent, does for free, and noted the second live demo flopped too. Muse already had its own cloud VM, connectors, and Mac computer use, and SpaceXAI’s Grok Bot shipped team agents the day before, while dot teams are “soon.” Even the name may be taken: the rumor is that SpaceXAI bought dot.com, and the domain now redirects to the Grok Bot download page. I checked the redirect myself, but the purchase is unconfirmed. (source: @OpenAI)

The White House Opened America.gov as One Front Door to the Government

The White House launched America.gov, which it describes as “a new national front door to the federal government,” one trusted place to find answers for whatever you need from government. According to Nick Sortor’s post, President Trump signed an executive order directing every federal agency to fold into it, creating one AI-powered front door for the whole U.S. government. I visited the site myself, and it’s a single chat experience that pulls government websites into one place, so renewing a passport or checking on a benefit starts in the same conversation. The White House post itself doesn’t explain how the AI works. (source: @WhiteHouse)

Anthropic Shipped Claude Opus 5.5 and Claude Sonnet 5.5

Anthropic released Claude Opus 5.5 on September 22 and Claude Sonnet 5.5 six days later. Anthropic says Opus 5.5 performs at the level of Claude Fable 5.1, its most capable model line, while costing about 40% less than Opus 5 on typical work. It’s now the default in the Claude app and Claude Code. Sonnet 5.5 costs the same as Sonnet 5 but runs more than 30% faster and needs fewer tokens per task. Most early praise was about writing: Opus 5.5 leads with the important information and follows your style rules. It’s also Anthropic’s first model since Dario Amodei’s essay on pacing the frontier, which Shikhar pointed out came out 11 days earlier. (source: @claudeai)

Meta Connect Put Muse on Glasses and a Keychain

At Meta Connect on September 23, Mark Zuckerberg announced that Muse, Meta’s personal agent that runs in its own cloud VM, now does voice and real-time video and keeps working while you talk, and it’s coming to Meta’s glasses. Muse is also getting an email address you can forward things to, and the Mac app got computer use. On hardware, Muse Charm is a keychain device for talking to your Muse, shipping in December, and Mark Gurman reports the new Meta VR Glasses cost $1,299 and weigh 100 grams by moving the battery and compute to an external pack. On September 29 Meta added Muse for Small Business, with connectors for Shopify, QuickBooks, Slack, Stripe, and more. (source: @finkd)

Amazon Blocked Muse While Shopify and PayPal Signed On

Amazon blocked Muse from shopping on users’ behalf, citing unauthorized access and privacy concerns. The same day, Shopify CEO Tobi Lütke announced agentic checkout with Shop Pay for Muse on every Shopify store, and PayPal followed a day later with Muse checkout across its merchants. Instacart is coming too: say “Taco Tuesday” and get a grocery cart from your usual store. Investor BuccoCapital called it the start of a “digital knife fight” between platforms that don’t want to be reduced to a back end for someone else’s agent. (source: @tobi)

Claude Found an Unknown Enzyme System in Viral DNA

Anthropic reported the first result from its new molecular biology lab. Claude found a previously unknown enzyme system in the DNA of bacteriophages, the viruses that infect bacteria, sitting next to a long repeating DNA array that looks somewhat like CRISPR, the basis of today’s gene editing. Nobody knows yet what it does, but the few known systems with these features can all cut, copy, and paste DNA. Claude read the literature and genome data, found the system, and proposed the experiments, and Anthropic’s scientists did all the lab work. Dario Amodei wrote that it’s work he’d have been proud of as a PhD student. (source: @AnthropicAI)

AI for Developers

GPT-6.1 Sol and the Rest of DevDay for Developers

Developers liked these launches more than dots. GPT-6.1 Sol nearly matches GPT-6 Astra on agentic coding and computer use at a fifth of the price: $2 input and $10 output per million tokens on OpenRouter, with cached input at $0.10. That came a week after GPT-6 Sol and GPT-6 Luna launched at half GPT-5.6’s prices. OpenAI also shipped Ultrafast, a paid speed tier at up to 300 tokens per second in Codex, a new Codex Cloud, a preview Agents API with computer use, MCP-based plugin extensions, and Sign in with ChatGPT, which Peter Yang called “a game changer” because users can bring their ChatGPT plan into your app. (source: @OpenAIDevs)

Claude Code Cloud Sessions Are Out of Preview

Cloud sessions keep Claude Code working with your laptop closed, and existing subscribers get a one-time credit to try them: $100 on Pro, $250 on Max. Anthropic also opened a plugin portal where you submit a remote MCP server or a GitHub bundle of MCP servers and skills, track review, and see installs; it says MCP usage across Claude products is up 110x this year. If you switched to Claude Opus 5.5, Anthropic’s cost guide says it thinks more per turn than Opus 5 at the same effort level, so start at medium instead of carrying over xhigh. (source: @ClaudeDevs)

Xiaomi’s MiMo-V2.6 Pro Is the Top Open-Weights Model

Xiaomi released MiMo-V2.6 Pro and Flash with open weights, training code, and RL environments. Artificial Analysis scores the Pro at 46 on its Intelligence Index, the highest of any open-weights model, a 1.02-trillion-parameter mixture-of-experts priced at $0.435 input and $0.87 output per million tokens. Xiaomi says it’s on par with Claude Opus 5 and GPT-5.6 Sol on most agent benchmarks, and OpenCode made Flash free for a week. For comparison, SpaceXAI’s closed Grok 4.7 also scored 46 that week at $2 input and $6 output. (source: @XiaomiMiMo)

Decision Models Are Now a Category

Two weeks after TypeSafe launched Jev, a model that returns typed answers with probabilities instead of text, OpenRouter counts six decision models on the market, two of them free. OpenAI joined at DevDay with a Decisions API powered by GPT-6 Luna, in limited preview: you define questions and possible answers to classify content, route requests, or choose an agent’s next action. OpenAI’s demo routes 10,000 customer requests at 150ms each versus 1.6s on the Responses API, which it calls about 10x faster. Google’s Gemma team published a one-command deploy for DiffusionGemma-Jev, a Jev-compatible endpoint on Cloud Run that runs about $3 an hour and drops to zero when idle. Trevin Chow posted the useful counterpoint: in a re-ranking pipeline with evals, Jev was about 50x cheaper and 25x faster, and he kept Opus anyway because its rankings were better. His read is that the value lies in new loops that cheap decisions make affordable. (source: @OpenRouter)

WebMCP Reached Stripe Checkout and Meta’s Glasses

WebMCP lets a website hand an AI agent a set of tools directly, so the agent doesn’t have to click through the page. Stripe rolled it out across Checkout on September 22, and its benchmark found agents used 42% fewer tokens when paying through WebMCP. On September 28, Stripe’s Jeff Weinstein counted new support from Meta wearables, Kernel’s browsers, Cloudflare Workers, and Shopify checkouts, and Cloudflare added it to Kitesurf, its Workers-based browser for agents. (source: @jeff_weinstein)

Cloudflare Wants to Be a Post-Quantum Certificate Authority

For Birthday Week, Cloudflare applied to become a certificate authority, twelve years after it launched Universal SSL, with an ACME-first design and Merkle Tree Certificates meant to hold up against quantum computers. The developer releases: Worker Previews give every Git branch its own URL, configuration, and state, Vinext 1.0 runs Next.js apps on Vite in production, and EmDash 1.0 is an open source CMS for Astro with sandboxed plugins. The new cf CLI also replaces wrangler environments with a programmatic cloudflare.config.ts. (source: @Cloudflare)

New Voice Models From ElevenLabs and Google

ElevenLabs released Eleven v4 and Eleven v4 Turbo, ranked first by Artificial Analysis. You direct delivery with inline tags or a plain description of how a line should sound, instant voice clones need 10 seconds of audio, and Turbo’s median latency is about 100 ms for live agents. Google shipped Gemini 3.8 Flash TTS and Flash-Lite TTS with 2,000+ ready-made voices across 100+ languages and cues like <laughs>. Alibaba’s Qwen-Audio-3.1 cut speech recognition prices by up to 95%. (source: @ElevenLabs)

Honorable Mentions

For everyone:

  • ChatGPT Space is a shared workspace of interactive pages with charts, checklists, and dashboards that ChatGPT or your dot can keep updated, rolling out to Pro, Business, and Enterprise. (source: @ChatGPT)
  • Grok Bot is in Teslas, SpaceXAI’s agent ordering things and managing email by voice, for SuperGrok Heavy subscribers for now. (source: @SawyerMerritt)
  • Googlebook is Google’s new laptop platform combining ChromeOS and Android, built around Gemini and available for preorder. (source: @Google)
  • Anthropic’s IPO prospectus shows a $42 billion net loss in 2025, per Polymarket, which also reported an $84.5 billion compute deal with SpaceX through 2029. (source: @PolymarketMoney)
  • OpenAI formed a math advisory group including Edward Witten and Timothy Gowers, after reports that an internal model resolved over 100 open problems. (source: @choblin29)
  • Bambu Lab R1 is the 3D printer maker’s first laser cutter, with camera alignment and auto-feed, starting at $2,499. (source: @ritwikpavan)

For developers:

Try This Weekend

For everyone:

  • Ask America.gov something you’d normally hunt across agency sites for, like a passport renewal
  • Give Claude Opus 5.5 your writing rules at the start of a chat and see whether it sticks to them
  • Snap a photo of something broken in your neighborhood and ask Muse to report it, the way Todd Saunders got Verizon to look at tangled wires
  • If you run a small business, connect one tool through Muse for Small Business and hand it a weekly chore
  • On ChatGPT Pro, create your first dot from the ChatGPT desktop app and give it an errand you’ve already given another agent

For developers:

  • Run /claude-api prompt-audit in Claude Code to strip anti-patterns from your CLAUDE.md and skills, per Lance Martin
  • Have Opus 5.5 review your effort settings against Anthropic’s cost guide
  • Swap gpt-6.1-sol into one long agent run on OpenRouter and compare the bill against GPT-6 Astra
  • Deploy djev-run on Cloud Run and point one classification step at it
  • Add WebMCP to Stripe Checkout and let an agent try to pay

Last week: the September 20 roundup led with TypeSafe’s Jev coming out of stealth.