Seventy to seventy-eight dollars a share. That's the number Anthropic put in front of public investors yesterday. And somehow it still reads like a discount. Welcome to Lantern AI Daily Briefing. It's Wednesday, November 11th, and we've got four stories: Anthropic's IPO terms, Google finally opening up Gemini 4 Carbon, a Senate vote that could change who pays for data-center power, and a one-point-six-trillion-parameter open model from, of all places, a food-delivery company. One tap on follow, and we'll be back in your ears before you know it.
The PoYo team's daily roundup, drawing on Anthropic's announcement and reporting from TechCrunch and CNBC:
At that price range, the 825 million shares imply an offering size of about $57.75 billion to $64.35 billion. TechCrunch reported that, at the top of the range and on a fully diluted basis, the company would be valued at about $1.404 trillion, above the $965 billion valuation reached in a May fundraising but below a target of about $2 trillion reported earlier.
This is the next step in the IPO story we've been following. On Monday we covered investor questions about whether a roadshow could even start this week. It started Tuesday. Anthropic is offering eight hundred twenty-five million Class A shares at seventy to seventy-eight dollars each, about eight hundred one million of them its own and the rest from existing holders. It has applied to list on the Nasdaq Global Select Market under the ticker ANT. And it's well under the Financial Times figure we cited last week, eighty to a hundred billion raised at up to one point seven five trillion. So the deal got smaller on the way to market. The part I'd read twice is the private placement. Twenty-six billion dollars of non-voting stock at the IPO price goes to AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC. TechCrunch says Google and Nvidia are in for ten billion each. Those are its suppliers buying in at the same price as everyone else. Per the filing as reported, run-rate revenue hit about eighty-one billion dollars at the end of October, up roughly ten billion in a month. CNBC adds a nine-month net loss of a hundred four billion, mostly fair-value changes on convertible notes and derivatives. And per TechCrunch, the founders' Class F share carries fifty point one percent of the vote. Class A buyers get exposure, not control. The bigger effect is on everyone else's valuations. OpenAI's one point two five trillion private round is about to have a public price to be compared against. CNBC, citing people familiar, says pricing is planned for November 18th, a Wednesday, with trading the next day and roughly a fifth of shares going to retail. The registration still isn't effective, so none of those dates is final.
Machine Heart, in a piece republished by 36Kr:
Compared with the previous Gemini 4 Argon, Carbon has a clearer positioning: it is not satisfied with quickly giving an answer, but is willing to spend more time and computing resources to truly complete a complex task.
Koray Kavukcuoglu's DeepMind post puts it the same way: Argon stays the default for most coding and knowledge work, and Carbon gets the hard, multi-step problems. We covered Carbon's announcement on October 29th, when it was limited to cybersecurity defenders in Google's Fairwind Program. The news on Tuesday is general availability, through the Gemini API in AI Studio and Gemini for Business for all paying developers, with a rollout to AI Ultra subscribers. Pricing is what matters for builders. Two dollars per million input tokens and ten output, the same as Argon, until November 30th. After that it doubles to four and twenty. Even then, that's half of Fable 5.2's list price, which we covered yesterday. And Google's own guidance matches what Anthropic said: use the cheaper model by default, and save the expensive one for hard problems. Google's benchmark claims: first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench, at less than half the cost per task of the next-best outside model. In Gray Swan testing, Google says automated cyberattack queries succeeded two point one seven percent of the time, down from nine point eight six for Argon. The feature I'd actually test is slash-swarm, in preview in Antigravity. A coordinator starts parallel workers in isolated git worktrees, and adversarial verifiers run tests and push back on every diff. That's the right design. What I want is the rate at which a swarm finishes without a human stepping in.
Nichola Groom, reporting for Reuters:
The US Senate voted to advance debate on a sweeping bipartisan energy permitting bill on Monday, handing a victory to Democratic Senators Martin Heinrich and Sheldon Whitehouse as they seek to pass the ambitious infrastructure legislation before Congress adjourns.
The vote was seventy-three to twenty-five to take up the Bipartisan American Affordability and Jobs Act. That's a procedural vote, so nothing has passed yet. Fifty-one Republicans, twenty-one Democrats and one independent voted yes. Heinrich and Whitehouse wrote it with Republicans Shelley Moore Capito and Mike Lee. It would speed up reviews for power plants, transmission lines, mines and pipelines, and it would require large data centers to pay more of the cost of the grid upgrades they need. That cost-shift is the part founders should notice. Right now, a lot of the grid upgrade bill for a new campus lands on ordinary ratepayers. The bill would move more of it onto the operator, with protections so households aren't left paying when a project fails. So the price of a megawatt goes up, but you might get it connected a lot sooner. The politics have shifted since Election Day. Whitehouse told Reuters that Democrats ran and won on affordability and data-center accountability. President Trump has endorsed the bill too. The no votes were twenty-three Democrats, Bernie Sanders and Republican Mike Rounds. Some environmental and tribal groups say changes made over the weekend don't go far enough. Next come amendment votes, final passage and the House. Watch the data-center provisions through the amendment process. If they survive, that's a direct change to the cost model for every gigawatt plan we've been talking about.
MLQ.ai, in its writeup of the release:
Meituan, the Chinese food delivery and e-commerce group, published the weights of LongCat-2.5 on Hugging Face under the MIT license on Nov. 9, about a week after introducing the model
The model card describes a mixture-of-experts model with one point six trillion total parameters and about forty-eight billion active per token, a million-token context, and text, image and video input. MIT is about as permissive as it gets, though MLQ notes it doesn't cover Meituan's trademarks or patents. There's already a discrepancy. Hugging Face's metadata lists one point eight trillion parameters, not one point six, and VentureBeat flagged the same gap. That matters for anyone sizing hardware to self-host. The model also carries a Preview label. On benchmarks, Meituan reports eighty-eight point four on Terminal Bench 2.1, up from seventy point eight for the previous LongCat. On DeepSWE it jumps from nine point seven to sixty-four point six. But all of it is self-reported, and AI Weekly says nobody had run independent tests at launch. And Meituan's comparisons are against older closed models like GPT-5.6 Sol and Claude Opus 5. Convenient. For context, last week we covered Kimi K3.1 Pro's open weights, and the twelve-point gap between open and closed models on Artificial Analysis. LongCat adds another trillion-scale Chinese release to that tier. Hosted launch pricing is thirty cents in and a dollar twenty out per million tokens. One caution from VentureBeat for anyone selling to government: check your contract terms before you put this into government-facing work. With Washington scrutinizing Chinese models, procurement risk may cost more than the inference.
Carbon's cyber scores and its jailbreak testing could be a whole episode on their own. That's the beat of our sibling show, AI Safety Daily: alignment, model evaluations, emerging risks and governance, every weekday, focused on what changed and what the evidence says. Look it up in your podcast app.
My homework for this week: run Carbon against Fable on the same hard ticket before November 30th, while the introductory price still applies. And we'll keep reading Anthropic's filing as the roadshow goes on. This has been Lantern AI Daily Briefing, part of the Lantern Podcasts family. We're back tomorrow, Thursday.