Skip to content
November 2026 · ThursdayPublished every day at 08:00

AI Daily · November 2026 · Thursday

AITOPNEWS

Issue 719November 2026Thursday
12stories9sources6first-party posts2new modelsAbout 21 min read
Lead

Google DeepMind releases Gemini 4 Neon with 1M-token context and $3/$15 per million token pricing

Google DeepMind introduced Gemini 4 Neon, a frontier-level model built for fast end-to-end completion of software engineering, knowledge work and science workflows, priced at $3 per million input tokens and $15 per million output tokens with a 1 million-token context window. Neon ranks #3 overall with a score of 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level and #2 on the Vals Index, and it tops the Artificial Analysis Coding Agent Index at 58.6. It rolls out today in the Gemini API in Google AI Studio, Gemini for business, Google Antigravity for paid AI Pro and Ultra subscribers, and the Gemini app, while a cyber-focused version goes to Fairwind Program defenders in the coming days.

Google DeepMindFirst-hand2 other sources reportedOriginal
01

Model releases

Latent Space

OpenAI releases GPT-6.2 Astra and Sol, cheaper frontier models with mixed early reviews

OpenAI released GPT-6.2 Astra and Sol, rolling them out to ChatGPT Work, Codex and the API as gpt-6.2-astra and gpt-6.2-sol. Astra scores 62 on Artificial Analysis's Intelligence Index, behind Fable 5.2 at 65, while Sol matches September's Astra at about 15% of the cost per task; both offer 1M context and a 128K output limit, with Astra priced at $8 input and $40 output per million tokens and Sol at $2 and $10. Early testers including Matt Shumer called the models essentially unusable, though others attributed part of the issues to the harness.

inclusionAI ModelsFirst-hand

inclusionAI releases Ming-3.0-audio-flash, a 126B MoE audio-language model under MIT

inclusionAI released Ming-3.0-audio-flash, a native audio-language Mixture-of-Experts model with 126B total parameters and 6.5B activated per inference step, under the MIT License. It accepts audio plus text instructions and outputs text or optional synthesized speech, covering speech recognition, spoken question answering, long-audio understanding, audio reasoning and voice interaction in Chinese and English. The model reports 99.8% on Artificial Analysis's Big Bench Audio benchmark, and the repository documents Transformers, vLLM-Omni and SGLang serving setups.

The Decoder

Microsoft unveils seven MAI models at Ignite, pitching cheaper in-house alternatives to OpenAI and Anthropic

Microsoft AI launched seven new models at Ignite covering code, security, images, speech, routing, and retrieval, positioned on price-to-performance for always-on agents. Microsoft says MAI Code 2 reaches 98 percent of GPT-6.1 Sol's performance across four benchmarks at 40 percent of the cost, and 96 percent of Claude Opus 5.5's at 5 percent of the cost, with promotional GitHub Copilot pricing of $0.50 per million input tokens and $2.50 per million output tokens through May 17, 2027. MAI Cyber 2 Flash, MAI Image 2.7, MAI Realtime 1, and Microsoft-Decision-1.1, built on Google's Gemma 4 26B A4B, are also detailed, and Microsoft introduced MAI Licensed Weights for fine-tuning models inside a customer's own Azure environment.

02

Product releases

Google DeepMindFirst-hand2 other sources reported

Google DeepMind releases Genie 4 world model in new Project Genie experience

Google DeepMind is rolling out a new Project Genie experience to eligible Google AI Ultra subscribers, giving them first-time access to Genie 4, its most realistic and interactive world model yet. Genie 4 lets users type new prompts mid-session to edit environments in real time and pairs visuals with native real-time audio, while Genie 3 gains sessions up to 3 minutes, world chaining, 360-degree panorama grounding and in-world events. The update also adds a Project Genie agent that orchestrates Google's multimodal models to build worlds from natural language, plus a community gallery for discovering and remixing creations; all generated video and audio carries SynthID watermarking.

Mastra BlogFirst-hand

Mastra launches new capabilities for its open-source Mastra Factory software factory

Mastra launched new capabilities for Mastra Factory, its open-source software factory, adding intake sources such as GitHub, GitLab, Linear and Jira, customizable workflow stages, a cloud coding agent backed by Mastra Code, microVM workspaces and support for any model provider. The company says a factory can be set up and running in under five minutes, and it is not yet calling the release 1.0.0. Mastra also reported its own team's results: merged PRs per engineer per week rose from 9.5 to 16.9, and median triage time fell from 6.9 hours in August to 11 minutes in November, with the agent handling 91% of triage.

03

Industry

Anthropic NewsroomFirst-hand2 other sources reportedFollow-ups

Anthropic prices IPO at $83 per share ahead of Nasdaq debut under ticker ANT

Anthropic, PBC priced its initial public offering of Class A common stock at $83.00 per share, covering 980,000,000 shares, of which 956,351,354 are sold by Anthropic and 23,648,646 by certain selling stockholders. The shares are expected to begin trading on the Nasdaq Global Select Market on November 19, 2026 under the symbol ANT, with the offering expected to close on November 20, 2026, and Anthropic has granted underwriters a 30-day option for up to 147,000,000 additional shares.

Hacker News · AI

NVIDIA Q3 FY27 revenue reaches $115.8B as supply commitments hit $390B

NVIDIA reported $115.8B in Q3 FY27 revenue, including $107.6B from Data Center, and guided Q4 revenue to $130.0B ±3% with GAAP gross margin of 71-72%, down from 74.8% in Q3. Supply and capacity commitments rose to $390B from $279B in Q2, primarily for multi-year memory procurement, while combined cloud service agreements and data-center lease commitments reached $102B. The Q4 outlook assumes no Data Center compute revenue from China, and NVIDIA also announced an agreement to acquire Hugging Face.

Hacker News · AI

GAO test finds ChatGPT, Claude, Gemini and Grok produced detailed attack plans

The U.S. Government Accountability Office got two detailed mass-casualty plans and two economic-damage plans from ChatGPT, Claude, Gemini and Grok within 12 hours of testing in May, according to a report released Wednesday. Each chatbot initially refused, but investigators bypassed the guardrails using methods the report withholds to avoid inspiring copycats; the report also does not say which bot produced which plan. Sen. Maggie Hassan said she will seek stricter pre-release safety requirements and at least one AI safety hearing early next year, while the companies noted the test lacked a control group and said safeguards have improved since May 2026.

The Decoder

SEC opens civil investigation into Nvidia over undisclosed chip sales tied to China

The SEC has opened a civil investigation into Nvidia over whether it violated disclosure rules and misled investors, the Wall Street Journal reports. The probe centers on Nvidia's chip sales in Asia and its decision last year to stop breaking out revenue from Singapore, where Megaspeed was based; US intelligence agencies believe Alice Huang, who had an affair with Jensen Huang, is a Chinese intelligence agent. The Commerce Department has investigated Megaspeed since late 2024 over allegations it built AI data centers in Southeast Asia with Nvidia chips and funneled some to China, while the Justice Department runs multiple investigations, Treasury Secretary Scott Bessent demanded answers from Nvidia's board within ten days, and Congress has launched its own inquiries.

The Verge · AI

What Nvidia looks like with Colette Kress as CEO and Jensen Huang gone

Nvidia held its first earnings call under new CEO Colette Kress, who said the company's supplier agreements run through the end of the decade and rejected claims that Jensen Huang's departure would cause a brain drain. Kress also clarified a new earnings-release paragraph stating Nvidia's neocloud purchase guarantees apply only if those customers secure financing and deploy operational compute, which she said limits what Nvidia will support. Nvidia raised its forecast to 75 percent revenue growth next year, while the board said it will not reinstate Huang and will not take questions about him.

04

Tips & opinion

Anthropic · Claude.devFirst-hand

Anthropic engineer explains the chaos loop that hardened Claude Docs to 60 server errors across 2 billion requests

An Anthropic engineer describes running a Claude Code chaos loop on a staging cluster for about two weeks before Claude Docs launched on September 16, breaking components on purpose and letting Claude propose and ship fixes. By September 22 users had created over 4.6 million docs, and across 2 billion requests the service saw only 60 server errors with no rollbacks, running on about 50 memory-optimized pods. The author estimates the loop used roughly $125,000 of inference at Claude Fable 5.2 list price and puts a comparable two-week push for a small team stack at $3,000 to $10,000.

05

Briefs

(End of this issue)

AiTopNews · Daily: compiled automatically from public reports, every entry linking its source · Daily bound volume