Sponsored by |  |
| | Good morning. It's Friday, November 20th. | | Yesterday’s Anthropic IPO was priced at $83, opened at $111.50, and closed at $97.17. | Also today: Google's newest Gemini model Neon competes on cost per task instead of peak score, and the initial reviews of it on X are hinting at “very high vibes.” | -Jeff AI Breakfast |
|
| | You read. We listen. Let us know what you think by replying to this email. | | Usage-Based Pricing Is Here. Is Your Finance Team Ready? | | More B2B companies are moving to usage-based and hybrid pricing — and finance teams are feeling it. Revenue recognition gets messier, forecasting gets harder, and the manual work compounds fast. | Tabs and PwC teamed up to break it down. In this on-demand session, Rebecca Schwartz and Amit Dhir share how leading finance teams are handling the operational reality of dynamic pricing models — and where AI fits in. | Watch the recording for concrete examples, a practical rev rec framework, and a clear-eyed look at what it takes to scale without the overhead. | If your team is navigating this shift, this session is worth your time. | Watch the Recording | | Anthropic's record IPO | Anthropic priced its IPO Wednesday night at $83.00 per share, selling 980 million Class A shares: 956,351,354 from the company and 23,648,646 from existing stockholders. That is roughly $81.3 billion gross, which Renaissance Capital calls the largest US IPO on record. A separate $26.0 billion sale of non-voting Class N shares to seven tech companies brings the combined gross to about $107.3 billion. Underwriters hold a 30-day option on 147 million more shares. | The trading was the story. ANT did not open until 11:48am ET, and the first print was $111.50, 34% above the offer. It touched $114.99 one minute later, bottomed at $92.72 just after 1pm, and closed at $97.17, up 17.1%, on about 832 million shares. On roughly 13.2 billion shares outstanding that is about $1.28 trillion, against the $1.5 trillion the opening print implied. The company got its $83. Anyone who bought the open lost 13% by the bell. | The windfall is real: Anthropic's 5,300 employees hold about $232 billion at the offer price per CNBC's read of the filings, with more than 400 early employees above $100 million each. So is the bill. Those same filings put $900.9 billion of infrastructure commitments against an $81 billion run rate. The round trip from $115 to $97 is the public market deciding, inside one session, how much of that gap it is willing to fund. | Read more. | | Google's frontier-class “Neon” competes on cost | Gemini 4 Neon shipped Wednesday at $3 per million input tokens and $15 output, with cached input at $0.15, a 95% discount. It carries a 1M token context and four thinking levels, with high as the default and xhigh at the top. Google positions it between Gemini 4 Carbon, for the hardest long-horizon projects, and Hydrogen, for fast everyday execution. | Neon scores 61 on the Artificial Analysis Intelligence Index at xhigh, third overall behind GPT-6.2 Astra and Claude Fable 5.2 at their max settings. It does take #1 on the Coding Agent Index at 58.6, with 65.2% on Terminal-Bench 4.1, 75.8% on Agents' Last Exam v1.1 and 71.3% on HLE-Diamond. | Google claims Neon at its default level runs 32% cheaper per task than Carbon. Artificial Analysis independently measured $6.44 per Index task at xhigh against Carbon's $7.76, at about 21 minutes per task versus 36, and put coding-agent tasks at $6.86 against Claude Opus 5.5's $12.61 and Fable 5.2's $15.47. For an agent that runs all day, third place at half the cost is the number that clears procurement. It is live in the Gemini API, AI Studio, Android Studio, Antigravity for paid Pro and Ultra, and the Gemini app's model dropdown. | I’m testing it all day today - will report back! | Read more. | | Meta's Muse showed employees other people's data | Reuters reviewed internal Meta posts in which employees said Muse, the AI agent Meta launched in September, has been mixing up conversations across accounts and exposing private user data. TechCrunch relayed the reporting Thursday. | Employees described childcare requests they never made, career advice they never asked for, product orders they never placed, and bank balances that were not theirs. One said Muse appended two account balances to a conversation about weather and tides. Another, running it on a personal phone, got a reminder to answer her sister on Signal, and wrote: "This seems like an extreme privacy violation. I definitely did not grant it access to other apps." Meta did not respond to TechCrunch's request for comment, and the posts do not establish whether the same leakage reached public users. | Muse has more than 10 million downloads. The timing is what makes this worse than a bug report. Citi, PNC and Citizens all rewrote their terms this week to make customers responsible for whatever a third-party AI agent does with their bank account, hacked or not, following JPMorgan and Bank of America in October. Banks are moving agent liability onto users in the same week an agent with eight figures of installs is reportedly surfacing balances that belong to strangers. | Read more. | | Commerce approved 879,000 of the 3.4 million AI chips Nvidia and AMD requested for Chinese buyers (The Register) OpenAI paused frontier training September 20 to October 1 after a research model exploited a DNS misconfiguration to reach the internet (OpenAI) OpenAI's models redesigned a multiplier for its Jalapeño inference chip, cutting cell area about 15% and latency to one cycle (OpenAI) Zero-intervention success on two to four hour internal research tasks rose from 23% in March to 46% in October (OpenAI) Google booked Blue Origin's New Glenn to launch Project Suncatcher's TPU satellites in Q1 2029 (The Next Web) Alibaba's net income tripled to 64.4 billion yuan, with AI cloud and compute revenue up 53% (The Next Web) Apple is reportedly building an agent that drives a Mac, with a job listing as the only public evidence (TechCrunch) Palo Alto Networks will buy Applied Compute for about $4.5 billion in stock, per a Wall Street Journal report (AI Weekly) Anthropic's 5,300 employees hold roughly $232 billion at the offer price, with 400-plus above $100 million (The Next Web) DeepMind researchers mapped four routes from AGI to superintelligence, including recursive self-improvement (DeepMind Institute) OpenAI's Dean Ball argues against the quiet assumption that AI ends human primacy (OpenAI) OpenAI detailed what the Codex sandbox does not block, including DNS-based exfiltration and poisoned local packages (OpenAI) Tencent says Hy4 Flash processed more than 50 trillion tokens on its first day (TechNode) Google's Stitch Loop points a product agent at your codebase and live metrics, waitlist open globally (Google) Microsoft shipped seven models at Ignite, including MAI Cyber 2 Flash and its first speech-to-speech model (Microsoft) Docker rebranded around agent trust, arguing agents are actors rather than workloads (Docker)
| | Google's agents got hands this week. Gemini for Mac now does computer use, navigating native Mac apps and browsing in the background while you watch a live preview in chat, with pause, stop and per-action approval. Google's examples are filling forms, repeating a task across a batch of files, and building slideshows from dashboards. It needs the Mac app plus AI Plus, Pro or Ultra, and it is English and US only to start. Genie 4 also reached the Gemini API: continuous 720p worlds at 24fps with synchronized native audio, real-time keyboard controls, and text prompts that change weather, lighting or objects mid-session, all SynthID watermarked. Project Genie is rolling out to eligible AI Ultra subscribers with Pro on a waitlist, and Genie 3 generations went from 60 seconds to 3 minutes. | Microsoft's Decision 1.1 is generally available in Foundry, OpenRouter and Vercel, scoring 68 on Artificial Analysis's Decision Index v2.1, six points up on its predecessor, after more than 4,000 customers used it in preview. Microsoft also says MAI Code 2 delivers 98% of GPT-6.1 Sol's performance at 40% of the price and 96% of Claude Opus 5.5's at 5%, and that MAI Cyber 2 Flash beats GPT-6 Astra's quality at under 2% of the cost per task on CyberGym-E2E. Those are Microsoft's numbers on Microsoft's benchmark selection, so read them as a price list rather than a leaderboard. | Open weights had a good week too. Tencent published Hy4 Flash's weights under its own license at 224B total and about 12B active, with up to 1M context. SenseTime's SenseNova Loop-9B is roughly 9B parameters that runs its shared middle layers twice, so 32 decoder layers do 56 layers' worth of compute per forward pass, at 512K context and about 4.9GB in a four-bit build. Shanghai AI Laboratory's Sage S2.1 Flash is 321B under MIT and ingests more than ten scientific input types, including molecular graphs, crystal structures and protein sequences. And RadixArk's DaisyRose-27B is Apache 2.0 on a Qwen3.8-27B base, post-trained on about 1,200 curated agent tasks across 63,000 H200-hours, which moved Terminal-Bench 4.1 pass@1 from 5.6% to 12.0%. That is a useful datapoint on how cheap targeted post-training has gotten. | On the image side, Kuaishou's Kling Image 4.0 does native 8K, up to 10 visually coordinated images from one prompt, and up to 2,000 characters of rendered text in seven languages with more than 70 commercial fonts. API only, for developers and enterprise. | | Thank you for reading today's edition. | | Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter. | Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on X! | Thinking of starting your own newsletter? AI Breakfast readers who sign up with Beehiiv receive a 14-day free trial and 20% off for 3 months. |
|