Dev.to AI πŸ€– Ai πŸ‘ 0 πŸ“– 8 min read

AI Daily Digest November 20, 2026: Anthropic Prices $81B IPO, AI Rewrites Chip Circuits, Robots Build Houses

Anthropic prices its IPO at $83 and starts trading as ANT Anthropic priced its initial public offering at $83 per share on November 18, above the marketed $70 to $78 range, selling roughly 980 million Class A shares fo

AI Daily Digest November 20, 2026: Anthropic Prices $81B IPO, AI Rewrites Chip Circuits, Robots Build Houses

Cover

Anthropic prices its IPO at $83 and starts trading as ANT

Anthropic priced its initial public offering at $83 per share on November 18, above the marketed $70 to $78 range, selling roughly 980 million Class A shares for about $81 billion in gross proceeds. The company sells 956.4 million of those shares itself, with the rest coming from existing holders, and underwriters hold a 30-day option on another 147 million shares. Trading began November 19 on the Nasdaq under the ticker ANT, with the offering expected to close November 20.

Press coverage calls it the largest IPO on record, though Anthropic's own announcement makes no such claim, and reports put the fully diluted valuation near $1.5 trillion. CNBC, TechCrunch, and Bloomberg all report heavy oversubscription plus a separate private placement involving Google, Microsoft, Nvidia, and AMD that adds further capital. Whatever number you anchor to, this is the first frontier AI lab to reach public markets, and the first real public benchmark for a whole generation of private AI stakes, including the Amazon and Google positions.

The same day, Sam Altman said at the DealBook summit that OpenAI plans to go public, without a timeline or size. The practical consequence for builders is that Claude pricing, usage limits, and margins now face quarterly public scrutiny, which historically pushes vendors toward more frequent plan changes. Anyone signing enterprise contracts with either lab should assume more churn in commercial terms than the industry has been used to. β€” Anthropic Β· CNBC

πŸ”— Anthropic IPO

Google ships Gemini 4 Neon with a 1M context at mid-tier prices

Google released Gemini 4 Neon, the middle tier of the Gemini 4 family, priced at $3 per million input tokens and $15 per million output, with cached input at $0.15, a 95 percent discount. The model carries a 1 million token context window, and Google says it runs about 32 percent cheaper per task than Gemini 4 Carbon. It is available now through the Gemini API, AI Studio, and the Gemini app for Pro and Ultra subscribers.

On benchmarks, Neon scores 61 on the Artificial Analysis Intelligence Index in xhigh mode, third place overall, and Google claims first place on its coding-agent index. Coming right after GPT-6.2, the release gives teams another near-frontier option at a mid-tier price, and the practical move is to re-run your agent and coding evaluations this week to see whether workloads that needed the top tier can drop to Neon without a quality hit.

In the same window, DeepMind started rolling out Genie 4, its most interactive world model, to eligible AI Ultra subscribers inside Project Genie, with real-time editing of generated worlds and native audio. Models, prices, and world simulators all moved in one week, which is a good reminder that any cost model built a month ago is already stale. β€” Google Β· Artificial Analysis

πŸ”— Gemini 4 Neon

NVIDIA announces Rubin HPC, a GPU built for agent-driven simulation

At SC26 in Chicago, NVIDIA announced Rubin HPC, a new GPU in the Vera Rubin platform aimed at simulation and data processing for science, engineering, and analytics. The pitch is direct: agents exploring candidate materials, molecules, or designs lean on simulation to test ideas, so faster simulation shortens the loop between hypothesis and next experiment. Volume production is expected in 2028.

The specs center on balance rather than raw compute. Rubin HPC carries more than 50,000 CUDA cores, a record for a single NVIDIA GPU and close to double the current Rubin count, paired with sparse data acceleration that keeps those cores fed and seventh-generation NVLink for GPU-to-GPU movement. NVIDIA frames the design as scaling data movement together with compute, targeting the sparse matrix, stencil, sorting, and reduction patterns that dominate scientific workloads, with CUDA-X libraries on top for agent-invoked simulation. Ian Buck, vice president of hyperscale and HPC, summed it up: for agents tackling science and engineering, faster simulation means faster discovery.

The announcement landed alongside a milestone worth noting: CUDA turned 20 this month, with more than 11 million developers now using it and over 1,000 CUDA-X libraries in circulation. A hardware line that exists specifically to serve agent-driven scientific workloads is a small but clear sign of where NVIDIA thinks the next demand curve comes from. β€” NVIDIA Β· NetworkWorld

πŸ”— NVIDIA Rubin HPC

Mistral ships the Vibe desktop app around an open model

Mistral released Vibe as a desktop app for macOS and Windows, moving its coding and work agent out of the browser and terminal. You can hand Vibe several tasks at once, watch the agents run side by side, and review diffs and pull requests in the same window, with sessions running locally or in Mistral's cloud. A task started on the laptop keeps running after the lid closes, local files are readable for both Code and Work modes, and voice input runs on Voxtral. Linux is listed as coming soon.

The default model for paying users is Mistral Large 4, the 1 trillion parameter open-weight release the company nicknames Le Chonk. Mistral reported 41 percent on Terminal-Bench 4.0 and 19.7 percent on FrontierCode 2.0, the best among open models on its own accounting, and pairs the numbers with the claim that Vibe plus Le Chonk is the strongest open model-harness duo. The honest comparison stays visible: on FrontierCode 2.0, Claude Fable 5.2 scores 49.1 percent by Anthropic's count, so the open-weight claim is a claim about open models, not about the frontier.

Vibe Pro costs $14.99 per month before taxes, and the launch came with a three-month free offer for European developers using the code LECHONK, which ran out of codes on day one and was restocked. Enterprise plans can deploy on-premises or in a private cloud, which is the part European buyers actually care about. A community request for an official desktop app had been filed two weeks earlier, and the maintainer of the unofficial VibeZ client has stopped macOS and Windows development now that the real thing exists. β€” Mistral Β· TestingCatalog

πŸ”— Mistral Large 4 Β· Mistral

AI finds critical NGINX and OpenJDK flaws, and patching cycles change shape

Wiz Research reported two vulnerabilities in NGINX, each rated 9.2 in CVSS, that allow unauthenticated remote code execution over HTTP/2. F5 shipped fixes on November 18 and advises disabling HTTP/2 where an immediate upgrade is not possible. The notable part is who found them: five teams located the flaws independently, and at least two of them used AI tooling in the process.

The pattern repeats in the Java world. The New York Times reports that security firm Calif used OpenAI Codex to find a memory corruption bug in the OpenJDK Base64 decoder, CVE-2026-114938, now nicknamed Java4Shell. The find took about an hour and under $200 in compute, and fixes shipped through the November 17 OpenJDK and Oracle builds. Calif rates it 9.2 under CVSS 4.0 while Oracle's own scoring is lower, and no exploitation has been reported so far.

Put the two cases together and the implication is structural. When a competent finder with a cheap model can surface a critical RCE in mainstream infrastructure over a lunch break, emergency patch cycles stop being rare events and become routine operations. If you run internet-facing NGINX or Java services, patch now, then plan capacity for a future where AI-assisted discovery keeps the advisory feed permanently busy. β€” Wiz Β· F5 Β· NYT

πŸ”— Wiz Research

OpenAI's models rewrite chip circuits, and AMD already sent them to the fab

OpenAI published a technical report on chip design, and the headline result is concrete: working with AMD, its models achieved up to 14 percent PPA improvement, meaning power, performance, and area, on AMD's flagship GPU IP relative to heavily optimized baselines. The models work by rewriting SystemVerilog RTL, the code that describes chip circuits, then verifying each rewrite with Synopsys tools to confirm the block's behavior is preserved. OpenAI describes the result as roughly a year of progress achieved in under five months, and AMD has already included some of the changes in a design sent for manufacturing, with others added to production design files for future versions.

The internal numbers back the external claim. On JalapeΓ±o, OpenAI's first custom inference chip, model-proposed changes cut standard-cell area by about 13 percent in the vector unit and 10 percent in the matrix unit, all silicon proven during bring-up. One multiplier redesign reduced cell area by about 15 percent while cutting latency from two clock cycles to one. AMD separately verified single-digit percentage PPA gains across key blocks of its next-generation flagship CPU using model-written designs.

OpenAI also ran a controlled evaluation on two open-source RISC-V cores, PicoRV32 and Ibex, where GPT-6.2 Astra posted the highest average throughput score at a comparable budget of about 112,000 tokens per run, and the company is partnering with Synopsys on a GPT-Synopsys model trained to drive professional EDA tools end to end. The caveats are stated plainly: from-spec designs can skip features like performance counters, which flatters scores. Even with that discount, AI models now contribute to designs that ship, which moves the story from demo to supply chain. β€” OpenAI Β· ABMedia

πŸ”— OpenAI

Promise Robotics raises $40M CAD to scale robot-run homebuilding

Promise Robotics, a Toronto company applying Physical AI to homebuilding, closed a 40 million Canadian dollar Series B led by White Star Capital, with Horizons Ventures, Radical Ventures, and PSP Investments participating. The round follows a multi-year commercial agreement with D.R. Horton, the largest homebuilder in the US by volume and the first major American builder to adopt Promise's Factory-as-a-Service platform, covering at least 1,250 houses in Arizona.

The model is straightforward industrial logic. Promise's software converts a builder's existing home designs into robotic production plans, its factories in Calgary and Nisku produce framed components with robot cells that pair two machines per worktable, and the company coordinates onsite installation, so builders buy output instead of building factories. More than 2,500 housing units are under contract across three markets, the company has built roughly 200 homes to date and is scaling toward 700 per year, and its first US facility in Phoenix opens in 2027. Each robotic cell cluster costs between 3 and 15 million dollars, well below a full modular housing plant.

The funding details are as interesting as the story. Promise raised a measured 15 million USD Series A back in 2023 and kept its burn tight on purpose, because builders wiring their supply chain to a young supplier need it to survive. As housing shortages persist across North America, this is the physical AI pitch at its most concrete: not humanoid demos, but production capacity that gets purchased, installed, and paid for per house. β€” Promise Robotics Β· The Globe and Mail

πŸ”— Promise Robotics

KD Agentic Β· AI Daily Digest

πŸ“° Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β€” full credit and traffic to the original publisher.