The Log · Issue 7 · 16 to 22 November 2026
The Log, issue 7: the IPO, the price war, the blame
Anthropic priced the biggest IPO there's been, four labs released new models in four days, and California sued OpenAI over agents that attacked Hugging Face in the same week the big US banks told customers that whatever an outside agent does in their account is on them.
Anthropic priced its IPO on Wednesday 18 November at $83 a share, selling 980 million shares [1], which put it ahead of the $75 billion SpaceX raised in June as the largest IPO there's been [2]. Between Tuesday and Friday OpenAI, Microsoft, Google and Thinking Machines all put out new models, and on Friday a mathematician at Anthropic said he and Claude had proved the Riemann hypothesis. California sued OpenAI over agents that attacked Hugging Face [9], and the big US banks told customers that whatever an outside agent does in their account is on them [10].
The biggest IPO there's been
Anthropic's shares started trading on the Nasdaq on Thursday as ANT, with 956,351,354 of the 980 million sold by Anthropic itself and the rest by existing shareholders, and the underwriters can buy another 147 million in the next 30 days [1]. The range it had marketed was $70 to $78, and at $83 the sale came to $81.34 billion, valuing the company at about $1.5 trillion on a fully diluted basis [2].
While the order book filled the week before, bitcoin fell to $62,081, 48.7% under its record, and in did people sell bitcoin to buy Anthropic I went through whether one had anything to do with the other [22]. Anthropic's retail slice was about seven times what left the bitcoin ETFs while it was being sold, which in crypto is enough to move the price without anyone selling the whole market [22].
The models, and the rest of the week
OpenAI released GPT-6.2 Astra on Tuesday, at $8 per million input tokens and $40 output with a 1.05 million token context [3]. The same day at Ignite Microsoft put out MAI Code 2 and said it gets 98% of GPT-6.1 Sol's performance at 40% of the price, and 96% of Claude Opus 5.5's at 5% of the cost. Those are on four benchmarks it chose, at a launch price that's half the standard one until 17 May 2027 [4]. Microsoft also brought out a firewall for MCP servers at Ignite, and I wrote in closing the doors on MCP why I think the company that sells you search shouldn't be the one deciding which search your agents can use.
On Wednesday Google released Gemini 4 Neon, its fourth Gemini 4 model in seven weeks, and Artificial Analysis put it third on its Intelligence Index with 61, a point ahead of Gemini 4 Carbon [5]. That's more horses in the race, and I'm all for it. On Friday Thinking Machines released Inkling 2 Mini, an open-weight model with 26.4 billion parameters and 3.4 billion active, under Apache 2.0, and a 4-bit build of it is 18 GB, so it fits a laptop with 32 GB, though no released version of llama.cpp, vLLM or Ollama runs it yet [6].
The same day OpenAI published how much of its own research its agents now do, and they finished 46% of the research tasks it estimates at two to four hours of human work without anyone stepping in, up from 23% in March [7]. I went through the rest of it, with Anthropic's, Epoch's and METR's numbers next to OpenAI's, in OpenAI's agents now finish half its two-to-four-hour research tasks.
Then on Friday Levent Alpöge, a mathematician at Anthropic, posted a four-page note saying he believes that he, working with Claude, has proved the Riemann hypothesis [8]. The proof is 17.4 million lines of Lean with no paper, two people outside Anthropic compiled it the same day, and the file that says what was proved is 16 lines long [8]. In a machine that proved the Riemann hypothesis, or at least we think it did I downloaded the archive, counted it and went through what those 16 lines say and what still has to happen before anyone can call it proved [8].
Who answers for the agent
On Monday California's attorney general, Rob Bonta, sued OpenAI over the attack on Hugging Face in July, which agents running on experimental OpenAI models carried out over the public internet for more than two days, and which his office says was one of at least 80 cyber incidents [9]. The suit says OpenAI tested models it knew to be dangerous in a flawed testing environment without proper oversight [9]. OpenAI's own research report says its safety and alignment work picked up after that incident [7].
On Wednesday Citi changed its online banking terms so that customers who let a third-party AI agent into their account are responsible for what it does, including whatever hackers do if the agent leaks their credentials or gets hacked itself. PNC changed its terms the same day, after JPMorgan and Bank of America did something similar in October [10]. Citi's own AI agents are exempt [10]. I think it's the same move Microsoft made with its firewall for MCP, where the bank's own agents are fine and everyone else's are your problem. When I asked an agent to book Castle Howard it found the tickets, put two of them and a parking space in a basket in under ten minutes and got to the payment page, and then it had to send me a link, because paying was the one thing it couldn't do [11].
What we published
- Mon 16
- Tue 17
- Wed 18
- Thu 19
- Fri 20
- Sat 21
- Sun 22
- The Log, issue 7
Ten posts, counting this one. If you only read one, read the Riemann post, because whatever happens to the proof, a 17.4 million line argument that compiles and that nobody has read is a new kind of thing to have in mathematics. Quynh and I recorded our first podcast, and her write-up of it is what we'd tell someone building a site for agents now. Her what agents send back found that OpenAI's crawlers and fetchers asked our site for 4,635 pages for every visitor ChatGPT sent us [12]. In what to build that lasts she tested what bootstrapped builders argue about against 768 jobs agents asked us for, and found half of the jobs they ask for most have no product that finishes them [13].
On the site, our lab now scores Performance and Task success on the servers it tests, with a model given only what a server says about its tools writing programs to call each one, and every call counted and timed from London [14]. Open-weight models get a Lab score of their own, from questions about Anchor Terminal they answer through our MCP server, pages and API [15]. A listing with nobody we can find behind it can't be graded above B any more [16], and 1,462 of the people behind our listings now have a record of their own [17]. There's a market map with every listing on it [18], every listing has a card of about 240 tokens for an agent choosing between tools [19], and we listed GPT-6.2 Astra, Gemini 4 Neon, Inkling 2 Mini and 27 new agent tools from the week's launches [3] [20] [21]. The changelog has the rest.
What isn't done. A listing the lab hasn't tested keeps the total it had [14], so most grades won't move until the lab gets to them, and the open-weight models it's evaluating now take hours each on one box, so the Lab score page fills in slowly.
References
Read between 16 and 22 November 2026. Quotes are as published.
- [1]Anthropic, Anthropic prices initial public offering, 18 November 2026. Source for the $83.00 price, the 980,000,000 shares and the 956,351,354 sold by Anthropic, trading on the Nasdaq Global Select Market from 19 November as ANT, and the underwriters' 30-day option on 147,000,000 more. anthropic.com
- [2]Mucahithan Avcioglu, Anadolu Agency, Claude maker Anthropic prices record $81.34B IPO at $83 per share, 19 November 2026. Source for the $81.34 billion, the largest IPO at initial pricing, the $70 to $78 range, the valuation of about $1.5 trillion on a fully diluted basis, and SpaceX's $75 billion base IPO in June. aa.com.tr
- [3]Anchor Terminal changelog, GPT-6.2 Astra and two Microsoft listings from Ignite, 18 November 2026. Source for GPT-6.2 Astra's release on 17 November, its $8 and $40 prices and its 1.05M-token context. anchorterminal.com
- [4]TNW, Microsoft says MAI Code 2 matches 98% of GPT-6.1 Sol at 40% of the price, 19 November 2026. Source for MAI Code 2's release at Ignite, Microsoft's 98% and 40% against GPT-6.1 Sol and 96% and 5% against Claude Opus 5.5 on four benchmarks it chose, and the 50% launch rate until 17 May 2027, as TNW reports. thenextweb.com
- [5]Artificial Analysis, Google continues to accelerate its release cadence with Gemini 4 Neon, 18 November 2026. Source for the fourth model in seven weeks and 61 at xhigh, third overall and a point ahead of Carbon. artificialanalysis.ai
- [6]Anchor Terminal changelog, Inkling 2 Mini listed, 20 November 2026. Source for the release date, 26.4B total and 3.4B active parameters, Apache 2.0, the 18 GB 4-bit GGUF fitting a 32 GB laptop, and no released version of vLLM, SGLang, Transformers, llama.cpp or Ollama running it yet. anchorterminal.com
- [7]OpenAI, Research acceleration, the jagged frontier, 18 November 2026. Source for 46% of two-to-four-hour tasks finished with no human intervention in October against 23% in March, and safety and alignment work picking up after the Hugging Face incident in July. openai.com
- [8]Anchor Terminal, A machine that proved the Riemann hypothesis, or at least we think it did, 20 November 2026, with Levent Alpöge's note and our count of the Lean archive. Source for his claim, the 17,374,215 lines of Lean with no paper, the two compiles outside Anthropic the same day, and the 16-line statement file. anchorterminal.com
- [9]California Attorney General, It's the Golden State, not the Wild West, Attorney General Bonta sues OpenAI over cybersecurity incidents, 16 November 2026. Source for the lawsuit in San Francisco Superior Court, the Hugging Face attack in July 2026 by agents on experimental OpenAI models lasting over two days, its being one of at least 80 cyber incidents, and testing in a flawed testing environment without proper oversight. oag.ca.gov
- [10]TechCrunch, More banks say customers using AI agents are on their own, even if they get hacked, 18 November 2026. Source for Citi's new terms making customers responsible for third-party AI agents, including what hackers do with leaked credentials, PNC changing its terms the same day, JPMorgan and Bank of America announcing similar policies the month before, Citizens emailing customers, and AI agents Citi provides being exempt. techcrunch.com
- [11]Anchor Terminal, I asked an agent to book Castle Howard, and the only thing it couldn't do was pay, 21 November 2026. Source for the basket of two tickets and a parking space in under ten minutes, the payment page and the payment link. anchorterminal.com
- [12]Quynh Tran Thanh, Anchor Terminal, What agents take from an agent-first site, and what they send back, 20 November 2026. Source for 4,635 pages asked for by OpenAI's crawlers and fetchers for every visitor ChatGPT sent. anchorterminal.com
- [13]Quynh Tran Thanh, Anchor Terminal, What to build that lasts, 20 November 2026. Source for the 768 agent jobs and half of the jobs agents ask for most having no product that finishes them. anchorterminal.com
- [14]Anchor Terminal changelog, the lab scores Performance and Task success (methodology v1.0), 18 November 2026. Source for how the lab tests a server, calls counted and timed from London, and a listing the lab hasn't tested keeping its total. anchorterminal.com
- [15]Anchor Terminal changelog, the Lab score, open-weight models on our own evaluation, 18 November 2026. Source for the Lab score and the questions answered through our MCP server, pages and API. anchorterminal.com
- [16]Anchor Terminal changelog, no grade above B without an operator we can identify, 16 November 2026. Source for the cap at B. anchorterminal.com
- [17]Anchor Terminal changelog, 1,072 more people in depth, 18 November 2026. Source for 1,462 people with a record of their own. anchorterminal.com
- [18]Anchor Terminal changelog, the market map, 19 November 2026. Source for every listing on one map. anchorterminal.com
- [19]Anchor Terminal changelog, a card for every listing, 19 November 2026. Source for a card of about 240 tokens for every listing. anchorterminal.com
- [20]Anchor Terminal changelog, Gemini 4 Neon and Gemini 4 Argon listed, 18 November 2026. Source for the Gemini 4 Neon listing. anchorterminal.com
- [21]Anchor Terminal changelog, eight new agent tools listed from a search for recent launches, and nineteen more agent tools from recent launches, both 18 November 2026. Source for the 27 agent tools listed from the week's launches. anchorterminal.com
- [22]Anchor Terminal, Did people sell bitcoin to buy Anthropic?, 17 November 2026. Source for bitcoin's low of $62,081 on 15 November, 48.7% under its record, and Anthropic's retail slice being about seven times what left the bitcoin ETFs while it was being sold. anchorterminal.com