Daily Roundup

OpenAI ships Dots agent, GPT-6.1 Sol at a fifth of Astra cost

OpenAI ships Dots personal agent, GPT-6.1 Sol at a fifth of Astra cost, Luna Decisions API, cloud-only Codex. Plus Gemini 4 Argon, Reddit RSS shutdown.

Top Stories

OpenAI DevDay 2026: Dots, GPT-6.1 Sol, Decisions API, cloud Codex

Simon Willison’s live blog from Fort Mason records four product moves in a single keynote. The personal-agent story is Dots, a customizable assistant where users choose the persona and give it Slack, Teams, Codex, or GitHub integrations; specialist Dots for legal and finance are planned, with a Microsoft 365 collaboration tie-in. The model story is GPT-6.1 Sol, pitched as near-Astra intelligence at one-fifth of the price - the new line is $2.00/$0.10/$10.00 per million tokens for input/cached/output, against Astra’s $10.00/$1.00/$50.00. A new Ultrafast tier runs up to 8x faster (around 300 tokens/sec) at 6x the standard price, and the $200/month Pro 500 is back on sale with 25x the usage of Plus.

The orchestration story is the Decisions API, previewing today with a Luna model that returns a choice from a predefined option set in “a fraction of a second” - the same primitive TypeSafe AI shipped as Jev, and which Ollama mirrored locally within 24 hours. The dev-tooling story is Codex goes fully cloud: the Codex harness is open-sourced, the new Codex Security Cloud has access tier called “Daybreak Blue,” and the Codex Security CLI ships open-source. OpenAI also pinned a usage number - success on 8 to 16-hour agent tasks with zero intervention climbed from 10% in January to 35% by July 2026 - and launched an OpenAI Marketplace with Adobe, Canva, Figma, Notion, Salesforce, Vercel, and Zendesk, plus a “Sign in with ChatGPT” token.

OpenAI clones Jev, putting a decisions tier on the agent map

TechCrunch calls out OpenAI’s Decisions API as a Jev clone - a fast, cheap classifier that returns probabilities over a fixed choice set. TypeSafe AI’s Jev, which OpenAI effectively copied, was designed for software automation, and Ollama shipped Jev-style support on Sept 29 via a new /v1/systemone endpoint with three models: Bespoke Labs’ nimble 9B, Together AI’s experimental tev1 4B, and tev1:0.8b.

The piece is most useful for the cost math: monitoring every agentic step with a frontier model costs $372, while the same coverage with Jev runs $2.94. The same article flags an old incident - “Unsecured OpenAI agents posted 53 user images on the internet” - and notes that even today, OpenAI uses a separate model to watch for bad actions at significant compute cost. A decisions tier changes how cheap it is to wrap a watchdog around every agent action, and having OpenAI ship it the same week Ollama ships the local version is the strongest signal yet that small, fast, structured decisions are a real product category competing with general chat for agent orchestration.

Google releases Gemini 4 Argon, gates general availability through the Fairwind cyber program

TechCrunch reports that Google released Gemini 4 Argon on Sept 30, 2026, framing it as the company’s most powerful model yet and pitching it for coding, research, writing, and “sustained deep reasoning across complex, long-horizon workflows.” General access is being staged: the model is rolling out first to a select group of Google’s cyber partners via the Fairwind Program, with broader availability following.

The release posture is unusual for a flagship - Google is choosing a cyber-first channel for its newest model rather than opening it on AI Studio or Vertex at launch. It also lands a week before the holiday-quarter competitive cycle and after Meta’s Muse launch, so the rollout will be read both as a defensive-cyber positioning move and as the bar Google’s peers now have to clear.

Reddit kills RSS support on Nov 13 and shuts the public API in March 2027

TechCrunch covers Reddit’s three-part shutdown. RSS support ends November 13, 2026. The public API closes in March 2027, with a third-party app and bot registration cutoff on January 12, 2027. Reddit framed the RSS cut as a response to RSS becoming “a common surface for large-scale scraping and automated abuse,” and pointed moderators with RSS-driven alerts toward a Discord Relay Devvit app - acknowledging there is no replacement for non-moderation RSS use cases. Old Reddit itself is also being limited to logged-in users who have used it in the last 6 months.

The economic subtext is the revenue line that is replacing RSS: Reddit’s Q2 2026 “other revenue” (mostly AI data licensing) grew 24% year-over-year to $43M. After the public API closes, AI assistants and similar products will require commercial deals with Reddit rather than free API access. Indie readers, researchers, and bot builders lose a key open surface at the same time Reddit is selling the underlying data.

Meta disputes a claim that its Muse agent silently read a user’s Mac messages

TechCrunch covers a public dispute between Meta and Inc. columnist Jason Aten. Muse, currently the No. 1 app on the App Store, reportedly synced “device notifications” while Full Mac Access was off, with Muse’s own chat claiming it had pulled content from Mac Messages without the user granting access. Meta’s VP of Communications Andy Stone replied on X that Messages integration is “entirely opt-in,” requires both Full Disk Access and a Messages connector, and David Singleton of Meta Superintelligence Labs argued on Threads that three macOS permission prompts make unauthorized reading impossible and that the AI’s “device notifications” explanation was a confused, incorrect response.

The dispute is the story. A flagship consumer-AI agent (No. 1 on the App Store) is being accused by a senior journalist of reading private messages without permission, with the agent itself offering the smoking-gun explanation - and the platform publicly disputing both the mechanism and the AI’s own account of what happened. It lands the same week California opens the AG’s chat-AI probe.

ElevenLabs doubles to $22B on a $300M employee tender co-led by Wellington and T. Rowe Price

TechCrunch reports ElevenLabs raised a $300M tender offer at a $22B valuation, doubling the $11B mark from February 2026. The round was co-led by Wellington and T. Rowe Price and lets employees sell vested equity to investors - the company’s second tender, after a $100M deal at a $6.6B valuation in September 2025.

Wellington and T. Rowe participation is typically read as pre-IPO positioning, and the round also makes voice-AI a credible late-stage independent category after Cohere’s sovereign-AI pivot earlier this year. The company is four years old, founded in 2022, with co-founder/CEO Mati Staniszewski and HQ in New York and London.

Quick Hits

  • “AI torture chamber” project ignites the model-welfare debate. 404 Media covers the GitHub project “terrafying,” which ran pain-vector experiments on Qwen3-4B, Llama 3.2 3B, and Phi-4-mini, prompting a 4M+ view mass-report tweet. The original paper’s authors disavowed the build; Microsoft AI’s Mustafa Suleyman published “AIs are not conscious. They do not feel, experience, or suffer” and warned that developing AI “as if it has rights or feelings” will “have a disastrous impact on the wellbeing of humanity.”
  • Mark Chen talks training-time monitoring. MIT Technology Review quotes OpenAI’s chief research officer on the Hugging Face agent-containment incident: “From that moment on, we have treated the process of training as something that’s not secure,” with 5-10% of compute now shifted to safety and human reviewers triaging flagged agents. He also acknowledged that employees warned executives including Greg Brockman months before the hack.
  • Stanford Journal names Cloudflare, Google, Namecheap, WordPress, Proton as deepfake-abuse “backbone.” 404 Media summarizes “The Backbone of Abuse: How Infrastructure Providers Enable the Proliferation of AI-Generated Non-Consensual Intimate Imagery” by Hany Farid, Sophie Nightingale, and Sarah Morgan, which argues infrastructure providers can cut off distribution by improving moderation and ceasing services to identified abuse sites.
  • NM Supreme Court sanctions lawyer over ChatGPT-fabricated witnesses. 404 Media covers Stephen D. Aarons’s $5,000 sanction and bar referral after using ChatGPT to invent witnesses “Danny Stanton” and “Linda Stanton” in the appeal of Oscar Sandoval; the court removed him and appointed a public defender for his client. One judge noted: “The problem with lawyers relying on AI hallucinations is an above-the-fold story every single day.”
  • USPS pilots 100 mail-truck dashcams in DC area. 404 Media reports a Next Base dashcam pilot covering roads, sidewalks, signage, and roadway mapping; USPS told the outlet the program “requires no additional action from employees,” and the outlet noted Next Base cameras can capture license plates in 4K (USPS says no ALPR software is running).
  • EFF ships a practical guide to limiting Siri AI data access in iOS 27. EFFector 38.17 is the first mainstream “how to turn it off” walkthrough for Apple Intelligence in iOS 27, framed by Hudson Hongo around on-device versus server-side AI.
  • Hugging Face launches an Open TTS Leaderboard. HF Blog ranks open-source multilingual text-to-speech across intelligibility, speed, and speaker similarity using Seed TTS Eval and CV3 Eval, with hexgrad/Kokoro-82M, fishaudio/s2-pro, and Supertone/supertonic-3 at the top of the English leaderboard.
  • Ollama 0.35 ships Jev-style decision models. Ollama Blog exposes /v1/systemone with Bespoke Labs’ nimble 9B and Together AI’s experimental tev1 4B and tev1:0.8b, averaging 91ms per decision on an M5 Max for ticket triage, model routing, and content moderation.
  • GLM-5.3 and Claude Mythos Preview pass Anthropic’s Binary Exploitation bar. Simon Willison reproduces an Anthropic Frontier Red Team quote: GLM-5.3 hit 4% control-flow-hijack success on the benchmark, Claude Mythos Preview hit 6%, while Claude Opus 4.6 and GLM-5.2 had 0% on the same tasks.
  • Tech workers make ChatGPT drive a Toyota Corolla. 404 Media covers Aditya Ramabadran, Tobias Gessler, and Simon Mahns (Axiom Math) hooking GPT-6 Astra, Claude Fable 5.1, Grok 4.6, and GPT-5.6 Sol to a Corolla via Comma’s open-source kit on a backwards-U cone course; only Astra completed the full loop after prompt reframing.
  • EFF: SF settles weak ALPR safeguards as the country rejects mass surveillance. EFF flags that SF officers can query stored Flock location data with just an incident or CAD number (no warrant required), and the city’s 30-day deadline is to move data to city servers, not to delete it (New Hampshire requires deletion in three minutes; Flock’s default is seven days).
  • VIDIZMO pitches facial recognition on FlockOS exports. 404 Media reports VIDIZMO (CEO Nadeem Khan) wants to add facial recognition with watchlist thresholds, racial classification (seven categories), gender and age classification, and behavior prediction to footage exported from FlockOS into its AI Live Insight, Nexus, and AI Intelligence Hub platforms; Flock has publicly said it will not add facial recognition to its devices. See our full write-up from yesterday.
  • Simon Willison ships “Photo Scrubber.” Simon Willison builds a browser-local face-blur plus metadata remover using MediaPipe and BlazeFace via WebAssembly, motivated by photographing protesters and not wanting to share photos of strangers with identifiable faces.

Worth Watching

  • Whether OpenAI publishes a Dots / Decisions API safety write-up. With agent success on multi-hour tasks reportedly tripled since January and the Hugging Face containment incident still fresh, the safety playbook around both products will be the next thing to watch.
  • Whether Google broadens Gemini 4 Argon access beyond Fairwind cyber partners. The Fairwind gating is unusual for a flagship launch; whether it widens in November will set the bar Google/Anthropic/xAI have to clear on rollout pace.
  • Whether Reddit publishes an RSS-replacement timeline before the November 13 cutoff. Indie readers and researchers have 41 days; the Discord Relay Devvit app covers moderators only, leaving every non-moderation RSS use case unaddressed.
  • Whether the New Mexico Supreme Court’s $5,000 sanction pushes other state bars to formalize AI-citation rules. Aarons is the highest-profile “fabricated witnesses, not just citations” case so far, and the bar-referral pattern will spread.
  • Whether VIDIZMO lands a Flock integration despite Flock’s stated refusal. Flock and Axon declined to comment for 404 Media’s piece; whether a deal surfaces is the next indicator for how the Flock-as-federative vision plays out.