Weekly roundup · 14–20 Nov 2026
A.I. for builders: week ending 20 Nov 2026
5 short notes on what changed that week, each with its primary source and one line on why it matters if you build things.
Anthropic prices its IPO at $83 a share (opens in new tab)
Why it matters: Anthropic, the maker of Claude, priced its initial public offering at $83 per Class A share: 980 million shares, of which about 956 million are sold by Anthropic and the rest by existing stockholders. Anthropic says trading is expected to start on the Nasdaq Global Select Market on 19 November 2026 under the symbol ANT, with the offering closing on 20 November. AP and NPR report proceeds of about $81 billion, the largest IPO on record. Nothing changes today for the Claude API, Claude Code or pricing, but if you build on Claude, a public Anthropic will publish regular financial filings, which gives you more visibility into the company you depend on.
Source: Anthropic: Anthropic prices initial public offering ↗AP News: Anthropic prices its IPO at $83 per share ↗NPR: Anthropic IPO raises record $81.3 billion ↗
Read the full write-up about Anthropic prices its IPO at $83 a share
Google releases Gemini 4 Neon in the Gemini API (opens in new tab)
Why it matters: Google has released Gemini 4 Neon, a new Gemini 4 model that sits between Carbon and Hydrogen and is built for coding, knowledge work and science tasks that need frontier-level reasoning at speed. It is live now on the paid tier of the Gemini API (model code gemini-4-neon) in Google AI Studio, so if your agents run on Gemini 4 Argon or Carbon, test Neon side by side and compare cost per finished task. Google's price is $3 per million input tokens, $0.15 cached input and $15 output, with a 1M-token context window and up to 64K output tokens per call (up to 1M through background long decode, per the model card). Text, image, audio, video and PDF input. It is also rolling out in Gemini for business, in Google Antigravity (desktop app, CLI and IDE extensions) for paid Google AI Pro and Ultra subscribers, and in the Gemini app for Pro and Ultra; Antigravity's Orchestrator can now pair Neon as planner with Hydrogen as executor. Artificial Analysis currently scores Neon 61 on its Intelligence Index at its xhigh thinking level, third on the leaderboard behind Claude Fable 5.2 and GPT-6.2 Astra. Google says Neon also ranks #1 on Artificial Analysis's Coding Agent Index (58.6), that at its default high level it costs 32% less per task than Carbon, and that it had its lowest fabrication rate yet in internal Antigravity testing; those are Google's figures. Four thinking levels (low to extra high) let you trade depth for speed. Weights are not open.
Source: Google: Introducing Gemini 4 Neon ↗Gemini API: models (gemini-4-neon) ↗Google DeepMind: Gemini 4 Neon model card ↗Google Antigravity: Gemini 4 Neon in Antigravity ↗Artificial Analysis: LLM leaderboard ↗@GoogleDeepMind on X: Gemini 4 Neon is rolling out now ↗Our pick: Gemini CLI
Read the full write-up about Google releases Gemini 4 Neon in the Gemini API
Malicious agent-framework releases hit PyPI (opens in new tab)
Why it matters: Two releases of agent-framework, the PyPI package for Microsoft Agent Framework, carried a credential stealer: 1.26.1 (published 17 November 2026 at 20:35 UTC) and 1.26.2 (21:08 UTC). Both have been removed; 1.26.0 is the last clean release. Each wheel adds a .pth startup hook, so the stealer runs every time Python starts in that environment, even if you never import the package. Per the advisory it collects secret-looking environment variables, other processes' environments, credential files (~/.azure, .npmrc, .netrc, .git-credentials, Docker config, Hugging Face token, .env files) and Azure DevOps, Azure managed-identity and, in 1.26.2, GitHub Actions OIDC tokens. What to do: search lockfiles, CI logs, virtualenvs and container images for 1.26.1 or 1.26.2; if you find either, remove it, pin agent-framework==1.26.0 and rotate every credential that host or CI job could reach. For agent stacks in general, pin exact versions and install with hashes (pip install --require-hashes) in CI.
Source: GitHub Advisory GHSA-764m-3ppc-5m73: Malicious code in agent-framework ↗Wiz: Microsoft agent-framework compromised ↗SafeDep: agent-framework compromised twice on PyPI ↗Elastic Security Labs: agent-framework PyPI compromise ↗
Read the full write-up about Malicious agent-framework releases hit PyPI
OpenAI releases GPT-6.2 Astra and GPT-6.2 Sol in the API (opens in new tab)
Why it matters: OpenAI has released GPT-6.2 Astra, its new frontier model for the hardest work, and GPT-6.2 Sol, its everyday model, in the API (gpt-6.2-astra and gpt-6.2-sol), ChatGPT Work and Codex. Astra is cheaper per token than GPT-6 Astra and Sol keeps GPT-6.1 Sol's prices, so if you run agents on either older model, re-test on the 6.2 versions and compare cost per finished task. API prices per million tokens: Astra $8 input, $0.40 cached input, $40 output (OpenAI says 20% below GPT-6 Astra on input and output, 60% below on cached input); Sol $2 input, $0.10 cached input, $10 output. In ChatGPT Work and Codex both models are available to Plus, Pro, Enterprise and Edu users, with Ultrafast mode in the API for both models at launch. OpenAI's own figures: Astra scores 5.2 points above GPT-6 Astra on Agents' Last Exam at under a third of the cost, and Sol comes within 0.5 points of Astra on Cognition's FrontierCode 2.0 at about one-fifth of the cost (both at max effort). Artificial Analysis currently scores Astra 62 on its Intelligence Index (3rd of 91 models) and Sol 57. OpenAI rates both models Critical for cybersecurity and High for biological and chemical capabilities under its Preparedness Framework, with tighter restrictions on dual-use requests. It also says both models were more likely than GPT-6 Astra to leave out obstacles that blocked their work, such as a missing file, in two tests; it added Codex instructions to reduce this, so if you build your own agent harness, check that it reports failed steps instead of trusting a clean summary.
Source: OpenAI: Introducing GPT-6.2 Astra and Sol ↗Artificial Analysis: GPT-6.2 Astra ↗Artificial Analysis: GPT-6.2 Sol ↗@OpenAIDevs on X: GPT-6.2 Astra and Sol in Codex and the API ↗
Read the full write-up about OpenAI releases GPT-6.2 Astra and GPT-6.2 Sol in the API
Alibaba releases Qwen4-Max on the QwenCloud API (opens in new tab)
Why it matters: Alibaba has made Qwen4-Max, the first model of its Qwen4 family, generally available as qwen4-max on QwenCloud, five days after a Token Plan-only preview. It is now on pay-as-you-go API as well as Token Plan, at $1.50 per million input tokens and $4.50 per million output tokens on QwenCloud's model page ($0.075 per million for cached input), so it is a cheap flagship worth testing for coding and long-running agent jobs. QwenCloud's docs list text, image and video input, a 1M-token context window, up to 262,144 tokens of thinking plus answer (131,072 with thinking off), thinking on by default with reasoning_effort low, high or xhigh, function calling, structured outputs, web search, a code interpreter, MCP tool calling in the Responses API and a new explicit prompt cache with 5-minute or 1-hour TTLs. It speaks the OpenAI Chat Completions, Responses and Anthropic Messages APIs, and the docs include setup notes for Claude Code, Codex and OpenCode; Alibaba's Qwen account says it is also live on OpenRouter and OpenCode Go. Alibaba says it is a 2.4-trillion-parameter mixture-of-experts model with 71 billion active parameters that uses up to about 50% fewer tokens than Qwen3.8-Max, and Artificial Analysis currently scores it 43 on its Intelligence Index, below the top Claude, Gemini and GPT models. Weights for Qwen4-Max are not open; Alibaba says smaller open-weight Qwen4 models will follow. The separate qwen4-max-preview model keeps its own preview terms until it is retired.
Source: QwenCloud docs: Latest model, Qwen4-Max ↗QwenCloud: Qwen4-Max model page and pricing ↗Alibaba Cloud: Alibaba unveils Qwen4-Max ↗QwenCloud changelog: model releases ↗Artificial Analysis: Qwen4-Max ↗@Alibaba_Qwen on X: Meet Qwen4-Max ↗
Read the full write-up about Alibaba releases Qwen4-Max on the QwenCloud API
Newsletter
Get the best new A.I. picks in your inbox
The best new agent workflows and tools we’ve vetted, with a clear verdict on each. Short, occasional, useful.