Anthropic put final numbers on its stock sale, mathematicians woke up to two competing claims on the most famous open problem in their field, and a small open model promised frontier coding scores on a laptop. Here is what held up and what still needs checking.
1. Anthropic Closes Its IPO With $93.5 Billion Raised
What happened. Anthropic announced the close of its initial public offering on Nov 20, 2026: 1.127 billion Class A shares at $83 each, including all 147 million extra shares the underwriters had an option to buy, for roughly $93.5 billion in gross proceeds, about $91.6 billion of which goes to the company. Alongside it, the company closed a private sale of about 313 million Class N shares at the same price to AMD, Broadcom, Google, Micron, Microsoft, NVIDIA and TSMC Partners, worth roughly $26 billion, bringing the combined total to about $119.5 billion.
The hype check. These are gross figures before underwriting discounts and fees, and a closing notice says nothing about how the stock trades from here. The private placement also deepens a circular pattern worth watching: several of the new shareholders are the same chip and cloud suppliers Anthropic pays for compute, and their Class N shares were not registered with the SEC.
Verdict: Supported. The figures come straight from the company’s closing announcement, and they follow our earlier note on how the shares opened on Nasdaq.
2. Two Claimed Riemann Hypothesis Proofs Land on the Same Day
What happened. Levent Alpöge, a mathematician at Anthropic, posted a short note claiming a proof of the Riemann hypothesis worked out with Claude, with the argument written in the Lean proof checker rather than as a conventional paper. Hours later, Stanford mathematician Ravi Vakil shared a draft by Alessandro Maria Masullo proposing an entirely human proof of the generalized Riemann hypothesis, noting that Brian Conrad had been reviewing it for months.
The hype check. A Lean check only confirms that the code proves the exact statement it was given, so experts still have to confirm that the formal statement matches the hypothesis itself, and the community is only starting that work. On the human draft, Vakil added an update within hours saying readers had flagged problems, some found with help from AI models, and he wrote that he makes no claim the strategy is correct. Neither result has been peer reviewed.
Verdict: Needs follow-up. Both claims are serious enough to track, and neither is settled.
3. Thinking Machines Ships Inkling 2 Mini, an Open Model for 32GB Macs
What happened. Thinking Machines released Inkling 2 Mini under the Apache 2.0 license: a mixture of experts model with 26.4 billion total and 3.4 billion active parameters whose middle layers run three times per token, which the model card says works out to about 5.4 billion effective active parameters. It takes text and images, handles up to 1 million tokens of context, fits in a checkpoint of about 20GB, and the company reports 79.2% on SWE-bench Verified. Ollama added support in version 0.40.17, on Apple silicon only for now.
The hype check. Every score is self-reported, the company says evaluation setups vary by model, and the card does not list sampling settings, seeds or reasoning effort for the main table. By its own numbers it trails Qwen3.8-27B on two instruction-following tests, and it needs at least 32GB of memory, so most base model laptops are out.
Verdict: Promising but unverified. The weights are open, so independent tests should arrive quickly.
4. Signal Moves to Block Phone Assistants From Reading Its Screen
What happened. Signal’s Meredith Whittaker said the app will now restrict operating system assistants from capturing its content. On iOS 27 Signal blocks screenshots inside the app, and on Android it detects when an assistant tries to read the screen and switches on its Screen security setting, which users can turn off. The post names Gemini on Android, Siri’s onscreen awareness on Apple devices, and Meta’s Muse agent as the kinds of tools it is guarding against.
The hype check. Signal admits the fix is partial: it does not cover macOS, where it can only suggest adding Signal to Siri’s excluded apps list, it does not work reliably on every Android device, and it does nothing about agents reading files already stored on a phone. The post also does not say how the change affects screen readers and other accessibility tools.
Verdict: News. This is a policy change by the app maker, and how Apple and Google respond is the part to watch.
5. METR Says Agent Time Horizons Now Double About Every 3.6 Months
What happened. The evaluation nonprofit METR published a capability update finding that the length of tasks frontier models can complete keeps doubling roughly every 3.6 months across its combined task suites, with current models reaching time horizons of up to 27 hours on a subset of tasks with checkable answers. Models usually matched or beat skilled humans on coding and optimization work, but METR found they showed worse judgment than human experts on open-ended tasks where feedback is slow.
The hype check. METR itself warns that its containerized tests may overstate real-world ability because similar environments are common in training, that many human time baselines are estimates, and that the trends exclude unreleased internal models. A faster doubling on the checkable subset may reflect a real speedup or something else, and METR says it is not sure which.
Verdict: Supported, as a measurement on METR’s own tasks. Extending the line into the future is a separate and much weaker claim.
Also on our radar
- OpenAI laid out its defenses against adversarial distillation, reporting that its internal red team’s success at extracting protected reasoning fell from 54.0% on GPT-6 Astra to 0.2% on GPT-6.2 Astra.
- Meta used its science summit to release UMA Nitro and an open Lean library called ATLAS that formalizes 26 open textbooks.
- Apple researchers posted Depth Pro 2, a single network that estimates sharp, metrically accurate depth for both images and video at full resolution.
- Digger released OpenComputer, an Apache 2.0 self-hosted coding agent that you summon from a Slack thread and that returns its work as a pull request.
What to watch
The Riemann story will move fastest: the key questions are whether outside mathematicians agree that the Lean statement in Alpöge’s release is the real hypothesis, and whether the problems flagged in Masullo’s draft can be repaired. Independent benchmark runs of Inkling 2 Mini should land within days now that the weights are public, and any reply from Apple or Google to Signal’s complaint would show whether app makers will get a proper way to opt out of assistant screen reading.
Sources
- Anthropic, Anthropic announces closing of initial public offering, including full exercise of underwriters’ option to purchase additional shares, and closing of concurrent private placement
- Levent Alpöge, remark (note on the Riemann hypothesis)
- Ravi Vakil, Hurried announcement: Alessandro Masullo GRH
- Thinking Machines, Introducing Inkling 2 Mini
- Thinking Machines, Inkling 2 Mini model card
- Ollama, Release v0.40.17
- Signal, Signal is Serious About Protecting Your Privacy
- METR, AI R&D-relevant model capabilities are improving rapidly and are getting harder to measure
- OpenAI, Combatting adversarial distillation
- Meta, Advancing Science Together: Highlights From the Open Source AI For Science Summit
- Hugging Face Papers, Depth Pro 2: Sharp and Temporally Consistent Metric Monocular Depth
- Digger, OpenComputer repository on GitHub
By Willa Ravenwire · The AI Bandwagon
