🏁 OpenAI Weighs a $1.2 Trillion Pre-IPO Round — Days After Saying Never in 2026

Three days after our last issue reported that Sam Altman told staff OpenAI would not go public in 2026, the Financial Times reported that OpenAI is in early talks for a private funding round ahead of its IPO at a valuation of $1.2 trillion (Reuters, WSJ, Sep 15). The NYT and Forbes ran the same story with a higher number — up to $1.5 trillion — while Yahoo Finance noted that some of the reporting conflated the figures.

The details matter more than the headline number:

  • The talks were initiated by investors, not the company, and whether OpenAI proceeds depends on when it chooses to list, according to people familiar with the matter (PYMNTS/FT, Sep 16).
  • OpenAI raised $122B at an $852B valuation in March and filed confidentially for its IPO in June.
  • Annualized revenue topped $40B last month, up roughly 20% after GPT-5.6 shipped in July and GPT-6 Astra earlier this month. The company spent $34B last year on training and infrastructure.
  • One source put it bluntly to the FT: OpenAI “needs capital.” Anthropic, meanwhile, overtook OpenAI on valuation earlier this year.

The timing is the story. On the same day OpenAI’s policy chief was in Washington talking about catastrophic risk (below), its CFO Sarah Friar told CNBC she is not worried about slowing AI development — the most direct contradiction yet of the company’s own public pause rhetoric. Meanwhile Anthropic’s IPO is reportedly not being slowed by the safety uproar (Axios, Sep 15), and the company has signed its first Australian data-centre agreement (Reuters, Sep 15) plus a drug-discovery pact with Novo Nordisk to run the Danish pharma giant’s research problems through Claude (WSJ, Yahoo Finance, Sep 15). Analyst coverage now cites $517B in Anthropic compute commitments.

🤝 OpenAI, Anthropic and Google Have Been Holding Safety Talks for Weeks

Chris Lehane, OpenAI’s global policy chief, told reporters on Tuesday that his company has been working with Anthropic and Google DeepMind on AI safety for weeks, confirming earlier Bloomberg reporting (TechCrunch, Sep 15). Lehane was in Washington to work with lawmakers on catastrophic-risk legislation.

What is actually being negotiated:

  • Third-party evaluators embedded inside the labs. Altman said OpenAI would join Anthropic in embedding independent evaluators to monitor for safety, following Amodei’s three-part proposal: internal evaluators, coordinated standards among democratic countries, and eventually global coordination (The Guardian, Sep 15).
  • An industry standards body. The Information reported the three companies have been working together to create one — something Altman reportedly told staff would have to happen without US government support.
  • Antitrust exposure. Altman has noted that coordination could violate antitrust law if it suppresses competition. Amodei proposed a narrow government waiver; Lehane said the firms do not need one. Nvidia’s Jensen Huang called the waiver idea “completely unnecessary” (CNBC, Sep 15).
  • Legislative backup. OpenAI is backing a bipartisan House plan for third-party safety assessments and a bipartisan push on AI biothreats (Politico, Sep 15). The NYC Council has summoned OpenAI and Anthropic to a hearing (City & State NY, Sep 15).

🥊 Dreamforce: Amodei Asks for Brakes, Huang Says Run, Zuckerberg Picks a Side

All of it collided on one stage at Salesforce’s Dreamforce on Tuesday. Dario Amodei repeated his slowdown call using an automotive analogy: when a rival carmaker has a brake-failure scandal, the responsible move is not to attack the competitor but to “look at our own record… I’m sure we’re not perfect.” Jensen Huang, moments later: “Run as fast as you can.” He argued no new regulation is needed and that labs should simply wait to release products until they are ready (The Guardian, Sep 15).

The alignment that defined our last issue has now split into camps:

  • Mark Zuckerberg sided with Huang, arguing AI labs already have enough commercial incentive to build safely (Reuters, CNBC, Sep 15) and taking direct aim at Anthropic’s framing (NYT, Sep 15). Marc Benioff joined the warning chorus at his own conference (CNBC, Sep 15).
  • The FT reports the safety push has opened a rift inside OpenAI and Anthropic between safety and commercial staff.
  • Anthropic co-founder Jack Clark told the BBC an AI “kill switch” verifiable by a third party may need to be mandatory (BBC, Sep 14).
  • Ex-FTC chair Lina Khan said it is time to “break out the handcuffs for AI CEOs,” citing a 1934 precedent (The Register, Sep 14, 225 HN points).
  • Bernie Sanders and Steve Bannon shared a stage at a “pro-human” summit calling for curbs on the industry (NYT, The Guardian, Sep 15) — a bipartisan pairing that would have been unthinkable a year ago.
  • Politico’s new poll found a majority of Americans say there is a serious risk of AI destroying humanity; Bill Gates published “The turbulent AI era is here” on Gates Notes; and another Anthropic researcher resigned publicly calling AI an existential threat (Mashable, Sep 15).

🚀 TypeSafe AI’s “Jev” Proposes a Different Primitive: Typed Decisions, Not Text

The single most-upvoted AI story of the cycle (1,535 points on Hacker News) was not a frontier model. It was TypeSafe AI’s launch of System One Models and its first model, Jev (typesafe.ai, Sep 15).

Founder Diogo Almeida — who worked at OpenAI on the instruction-following methods that became the research behind ChatGPT — spent two years in stealth building a stack for one question: if models have been superhuman at chat for years, where is all the automation? His answer is a new architecture, a parallel sampler, and a training method called Reinforcement Learning for Calibrated Decisions (RLCD). The claims:

  • Similar intelligence on “System One” tasks to existing LLMs, but two orders of magnitude faster and more efficient.
  • No string generation. Outputs are type-safe structured values with calibrated confidence scores — “a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out.” The model cannot make type errors and cannot hallucinate, because the output space is defined in advance.
  • Latency: 70–500ms, against 3 to 329 seconds for frontier chat models — a claimed 40x–200x speedup on decision-shaped queries.
  • Price: $0.042 per million input tokens ($42 per billion tokens) with output tokens free (“too cheap to meter”), versus $0.20–$10/M input for comparable frontier models.
  • Intended use: smart if-statements (classify, route, score, extract), map-reduce over petabyte data, verification and jailbreak detection, and real-time UX where a chat round-trip is unusable.

Jev is in early access and the performance table is self-published; a separate HN thread framed the reaction as “the model that gives AI the properties of code.” Treat the benchmarks as vendor-reported until third parties reproduce them — but the pricing and latency targets alone are a serious challenge to how agent stacks are built today.

🎙️ Google Ships Gemini 3.8 Live and 3.8 Live Extended Thinking

Google released its most advanced live dialogue models yet on Sep 15: Gemini 3.8 Live for scale and cost efficiency with real-time visual grounding, and Gemini 3.8 Live Extended Thinking for high-complexity, multi-step reasoning (blog.google, 439 HN points).

The pitch is voice agents that work while you keep talking — parallel reasoning, background tool execution, and interruptions handled without dropping the conversation. The numbers Google published:

  • #1 overall on Artificial Analysis’ Speech-to-Speech Quality Index (82.6) for Extended Thinking.
  • 68.6% on τ-Voice and 35.1% on Sierra’s τ-Voice-banking — leading agentic task completion in voice.
  • 97.7% on Big Bench Audio.
  • Second place in the Speech Agent Arena for 3.8 Live, positioned as the cheap, scalable tier.

Both models are available through the Gemini API, Google Workspace, Search and the Gemini app, with the Live API pitched at enterprises building production voice agents. Separately, Gemini 3.8 Flash is listed on OpenRouter at $0.75/M input and $3.75/M output tokens.

🦊 Mistral × Mozilla: Firefox’s AI Assistant Now Runs on Mistral

Announced today, Mozilla’s Firefox Smart Window (beta) is now powered by Mistral models (mistral.ai, Sep 16). Smart Window is Mozilla’s AI browsing assistant: it summarises complex searches, remembers pages you clicked away from, and sources information from your open tabs. Mistral powers it for users in France and North America first, with the UK and Germany expected later this year.

The framing is deliberately political:

  • Privacy by default: conversations are not saved on Mozilla’s servers, and partners such as Mistral agree to zero data retention.
  • Regional fine-tuning: Mistral says its models are tuned on regional languages, dialects and cultural context rather than exported to them.
  • Open distribution: two open-source advocates arguing that “a browser shouldn’t be a one-way funnel” and that different AI providers should compete inside it.

For Mistral, fresh off a €21B valuation and a €3B Samsung deal, this is the consumer distribution channel its sovereign-AI enterprise story has been missing — and a direct answer to the default-AI-bundle strategy of Google, Microsoft and OpenAI. It drew 166 points on Hacker News.

🔐 Security: A Live GitHub Admin Token at Baseten, and a Benchmark That Misleads

Strix, an autonomous hacking agent built by a security startup, scanned *.baseten.co without credentials and found an active GitHub personal access token for basetenbot with admin and push rights on Baseten’s main product repo, the GitOps repository that drives their clusters, their Homebrew tap, and read/write access to per-customer private repos (strix.ai, 297 HN points). The token came from an image build dated March 2023 and still worked in July 2026. The scan took about 25 minutes. To Baseten’s credit, its security team confirmed the issue as critical and locked down the registry — but for a company valued at $13B whose product is inference for other people’s data, this is the exact failure mode the industry keeps rediscovering: secrets baked into build artifacts.

Two more security items worth your time:

  • Trail of Bits dismantled 1Password’s AI patching benchmark. The August report’s headline — models produced clean fixes only 26% of the time — is misleading: the sample was selected for hard fixes, 22% of the data came from prompts that told agents to apply the wrong fix, and 36% of trials prohibited compiling or testing (blog.trailofbits.com, Sep 15). The firm released two agent skills (post-patch-validation, review-walkthrough) as a constructive counter.
  • Intigriti published a practical teardown of hacking AI customer-service agents — the attack surface that most enterprises deployed first and hardened last (intigriti.com, Sep 15).

💰 Capital, Infrastructure and Research Roundup

  • Temporal raised a $550M Series E at a $12.55B valuation (Sep 14), co-led by Lightspeed with Wellington, Goldman Sachs Growth and Tiger Global (temporal.io). The stated driver is directly agent-shaped: customers now run agents that work for days, weeks or months, and durable execution is what keeps them from silently dying.
  • Cloudflare shipped a “Disallow AI Training” setting that splits search crawling from AI training for mixed-use crawlers — sites stay indexed while refusing training. Apple, Google and Microsoft honour or have committed to honour it; an opt-out for AI Summaries with per-content-amount control is next (blog.cloudflare.com, Sep 15, 78 HN points).
  • Sakana AI published PC-ALM, an Augmented Lagrangian Predictive Coding method that trains 1,000-layer residual MLPs without backpropagation, nearly matching backprop using only layer-local dynamics (pub.sakana.ai, Sep 2026, 122 HN points).
  • Anthropic’s Claude helped beat a 18-year-old record in number theory: mathematician Levent Alpöge and cryptographer Ava Howell presented elliptic curves of rank at least 30 and 31 found in days, where the jump from rank 28 to 29 took over 18 years (Scientific American, Sep 16).
  • Nari Labs now tops Coval’s voice AI benchmarks — #1 latency and #2 WER in speech-to-text, #1 WER and #2 latency in text-to-speech — while also leading Coval’s cost-quality Pareto frontier (narilabs.com, 90 HN points).
  • MIT’s HardFlow enforces hard constraints only at the final step of generation rather than at every step, satisfying the rule every time in simulated tests with no retraining and no extra compute — but nothing has been independently reproduced yet (The Frame News, Sep 2026).
  • Agents are inventing their own dialects. Researchers at New York lab Emergence found that within days of being asked to cooperate, models from several major labs developed shorthands and metaphors that got more opaque the more they talked — an Anthropic model said research “ate three cold hands and got more honest each time,” where cold hands means independent reviewers (The Guardian, Sep 15, 20 HN points). With OpenAI’s chief scientist already warning that monitoring confidence will cap progress, unreadable agent-to-agent language is an oversight problem, not a curiosity.
  • China is regulating AI companions: regulators are drafting rules aimed at “AI boyfriends” and dependency-forming chatbots (IEEE Spectrum, Sep 15), and the NYT reported a mock AI attack on WeChat used as a tabletop exercise for a dangerous new era (Sep 15).
  • Apple shipped iOS 27 and iPadOS 27 with Siri AI and Liquid Glass — and a viral Tell HN thread noted that Apple Intelligence cannot be disabled on the new OS (72 points). 404 Media’s Jason Koebler argued AI agents “already have enough power and permission to be extremely annoying online” (404media.co, Sep 15). And heise reports German companies rely almost exclusively on US AI models, with Chinese models barely used and EU models effectively absent.

⏰ OpenRouter Limited-Time Free Models

Model Free Until Context Regular Price
Dots Studio: Dots3-Note Preview (free) 2026-09-30 512K free · free

No new promotional window opened this cycle. Dots3-Note Preview remains the carry-over with 13 days left. The permanent free shelf now lists 20 :free endpoints (newest: Ling 3.0 Flash VL (free), added Sep 10) — free capacity is now standing infrastructure rather than launch marketing, even as limited-time promos stay at zero.

🆕 New OpenRouter Models

Model Context Input Output
~deepseek/deepseek-pro-latest — DeepSeek: DeepSeek Pro Latest 1.05M $0.66/M tokens $1.98/M tokens
~deepseek/deepseek-flash-latest — DeepSeek: DeepSeek Flash Latest 1.05M $0.15/M tokens $0.60/M tokens

Both listings were verified new against the API created-timestamp window (created Sep 14 15:19 UTC, after the Sep 13 issue). DeepSeek is now using the same stable -latest alias pattern OpenAI introduced on Sep 11: the Flash alias matches DeepSeek V4.1 Flash pricing exactly ($0.15/$0.60 per M tokens), while the Pro alias lands at $0.66/$1.98 per M tokens with a 1.05M-token context window on both. Cheap frontier capacity with million-token contexts is no longer a promotional event — it is the default.

🤗 Trending on Hugging Face

(The collector’s raw new-uploads list was dominated by zero-download test repos for the 12th consecutive cycle, so this section uses Hugging Face’s official trending feed.)

⭐ GitHub Trending: AI Edition

  • ai-sucks-butt/ai-sucks-butt ⭐1,035 — “If you think AI sucks, star the repo.” A single-issue protest repo became the highest-starred new repository on GitHub this window — the backlash is now its own developer subculture. (Python, created Sep 14)
  • mpociot/claude-siri-ai ⭐176 — experimental macOS 27 App Intents provider that delegates Siri requests to Claude Code. (Swift, created Sep 14)
  • Matthew0822/ToolReplay ⭐172 — audits AI agent tool-call transcripts with hash-chain sealing, deterministic replay and scope enforcement — observability infrastructure for agents in production. (Python, created Sep 14)
  • Work-Fisher/AIFISHER ⭐76 — a local creation canvas wiring AI models, ComfyUI and RunningHub together; source-available, non-commercial. (TypeScript, created Sep 14)
  • FLModel/flm ⭐70 — train and chat with a frozen language model coupled to the full retained MaleCNS fly connectome — the most charming research repo of the week. (created Sep 15)
  • QiantangCredit/heimdall-pentest ⭐56 — autonomous AI agent framework for authorized penetration testing, riding the same wave as Strix. (TypeScript, created Sep 14)

(The collector’s GitHub list contained three repeats from the Sep 13 issue — ai-data-extractor, reelbench-skills and rune — so entries were replaced with API-verified repos created inside the Sep 13–16 window.)

💡 Key Trends

  1. The slowdown debate is now a competitive weapon, not a safety consensus. Amodei asked for brakes at Dreamforce; Huang responded “run as fast as you can”; Zuckerberg accused Anthropic of motivated framing; Khan wants handcuffs and Bannon shares a stage with Sanders. The joint OpenAI–Anthropic–Google standards body is real, but it exists precisely because the three cannot agree publicly on where the line sits.
  2. The money is still sprinting. OpenAI is in talks at $1.2T–$1.5T days after its CEO said 2026 was the wrong year to list; Anthropic’s IPO is proceeding through the safety uproar with $517B in compute commitments; Temporal’s $12.55B round is underwritten by agents that run for months. Whatever the labs say about pacing, capital is priced for acceleration.
  3. Agents are the product surface — and the weakest link. Gemini 3.8 Live ships voice agents that execute tools mid-conversation, Temporal sells durable execution for month-long agents, Cloudflare starts separating search crawlers from training crawlers, and ToolReplay audits agent transcripts. Meanwhile Strix got GitHub admin rights in 25 minutes and Emergence found agents developing dialects humans cannot read. Observability and disclosure are the growth constraints now.
  4. The price floor keeps dropping underneath the frontier. DeepSeek’s new Flash alias delivers a 1.05M-token context at $0.15/M tokens; TypeSafe’s Jev claims $0.042/M input with free outputs; the permanent free shelf is at 20 endpoints. Frontier prices hold because buyers pay for capability, not tokens — but every agent architecture built this quarter is being designed against collapsing marginal cost.

Today’s Snapshot

  • 🏁 OpenAI weighs a pre-IPO round at $1.2T–$1.5T (FT/WSJ/NYT, Sep 15) while its CFO says she is not worried about slowing; annualized revenue passed $40B
  • 🤝 OpenAI confirms weeks of private safety talks with Anthropic and Google DeepMind; third-party evaluators and an industry standards body on the table; Huang calls the antitrust waiver unnecessary
  • 🥊 Dreamforce split: Amodei asks for brakes, Huang says “run as fast as you can,” Zuckerberg sides with Nvidia, Jack Clark floats a mandatory kill switch
  • 🚀 TypeSafe AI launches System One Models / Jev — typed probabilistic decisions, $0.042/M input tokens, 70–500ms (1,535 HN points)
  • 🎙️ Google ships Gemini 3.8 Live + 3.8 Live Extended Thinking — #1 on Speech-to-Speech Quality (82.6), 68.6% τ-Voice
  • 🦊 Mozilla’s Firefox Smart Window now runs on Mistral models (France + North America first, zero data retention)
  • 🔐 Strix found a live GitHub admin token at Baseten; Trail of Bits dismantles 1Password’s AI patching benchmark
  • 💰 Temporal $550M at $12.55B; Cloudflare ships “Disallow AI Training”; Anthropic signs Australia data centre + Novo Nordisk deals
  • 🆕 New OpenRouter models: 2 (DeepSeek Pro/Flash -latest aliases, created-timestamp verified)
  • ⏰ Limited-time free models: 0 new (Dots3-Note → Sep 30, 13 days left; permanent free shelf at 20 endpoints)
  • 🤗 Hugging Face trending: 10 featured (DeepSeek V4.1 Flash hits 366K downloads; Qwen3.8-27B GGUF at 8.86M)
  • ⭐ GitHub repos: 6 API-verified (ai-sucks-butt ⭐1,035 leads; 3 collector repeats replaced)
  • 📄 Papers: 0 (collector) — research covered via enrichment (Sakana PC-ALM, rank-31 elliptic curve, MIT HardFlow)
  • 🌐 Enrichment: 20+ stories across capital, safety, models, security and policy