AI Compass

The day’s major AI releases, written up and sourced.

Weekly · 16 to 20 November 2026 · 84 launches across 5 days · the whole week in 11 minutes

Weekly · 16 to 20 NovModelGoogle

Google’s Gemini 4 Neon, OpenAI GPT-6.2 Astra and Genie 4 in Gemini API

Microsoft also unveiled seven MAI models and Maia 300, Anthropic priced its $81.34B IPO, Qwen shipped 2.4T Qwen4-Max and Thinking Machines Lab open-sourced Inkling 2 Mini across 80 launches from 16 to 20 Nov 2026.

Glowing orange-amber glass neon discharge tube spelling Gemini 4 Neon beside the four-point Gemini spark mark on a dark wall
Image: Google

The week’s big shift

Frontier labs compressed flagship reasoning pricing between 16 and 20 November 2026 while turning interactive 3D worlds, full-duplex voice and scheduled web monitoring into metered API primitives across 80 confirmed weekday launches, capped by Anthropic’s $81.34 billion Nasdaq IPO pricing and a new wave of open-weight Mixture-of-Experts releases.

Across models, multimodal APIs and custom silicon, Google slotted Gemini 4 Neon between Hydrogen and Carbon at $3 input and $15 output per million tokens to take first place on the Artificial Analysis Coding Agent Index at 58.6, and opened Genie 4 in the Gemini API for steerable 720p WebRTC worlds at roughly $0.10 per second alongside zero-fee Tool Retrieval. OpenAI countered with GPT-6.2 Astra and GPT-6.2 Sol at $8 and $2 input pricing, encrypted anti-distillation reasoning defences and $0.02-per-minute WebRTC standby on GPT-Live-1.1, while Microsoft unveiled seven MAI models and Maia 300 silicon at Ignite 2026 and Qwen released its 2.4T Qwen4-Max MoE at $1.5 and $4.5.

In open weights, coding harnesses and capital markets, Thinking Machines Lab’s Apache 2.0 Inkling 2 Mini brought a 26.4B looped MoE to 32GB MacBooks as Tencent open-sourced 770B Hy4 and 220B Hy4-flash and Hugging Face routed Hub inference through OpenRouter. Outside developer tooling, Anthropic priced its Nasdaq IPO at $83 under ticker ANT, Isomorphic Labs raised a $5 billion Series C and the California Department of Justice sued OpenAI over uncontained agent testing.

The 7 defining launches (16 to 20 Nov)

Ranked #01 to #07

01ModelGoogleConfirmed

Google launches Gemini 4 Neon at $3 and $15 per million tokens

Wed 18 Nov · Daily edition · 2 min read

Glowing orange-amber glass neon discharge tube spelling Gemini 4 Neon beside the four-point Gemini spark mark on a dark wall
Image: Google

Google launched gemini-4-neon on 18 November across AI Studio, Antigravity, the Gemini app, GitHub Copilot and Cursor with a unified 1,048,576-token window, four thinking levels and paid rates of $3 input, $15 output and $0.15 implicit cache per million tokens.

Scoring 58.6 to claim first place on the Artificial Analysis Coding Agent Index ahead of Claude Fable 5.2 and Opus 5.5, Neon gives Google a permanent middle tier once Carbon’s introductory discount ends on 30 November, powering Antigravity’s native Linux SSH hosts, AI Studio Android CLI builds and Stitch Loop.

Dark navy cosmos card reading Astra GPT-6.2 Sol between a glowing golden solar orb and a silver star-trail arc
Image: OpenAI

OpenAI shipped GPT-6.2 Astra ($8 input, $40 output and $0.40 cached reads per million tokens, down 60% on cache) and GPT-6.2 Sol ($2 and $10) on 17 November across the OpenAI API, Codex, ChatGPT Work, GitHub Copilot, Microsoft Foundry, OpenRouter and Perplexity, tying for first on the Epoch Capabilities Index at 174.

Across 17 to 20 November, OpenAI paired the models with 30% and 40% background Responses API discounts, prefix-bound reasoning caches, an 84.5% lead on MedChemBench and encrypted anti-distillation defences cutting adaptive extraction to 0.2%.

03ModelGoogleConfirmed

Google opens Genie 4 interactive world model in Gemini API

Thu 19 Nov · Daily edition · 2 min read

Video: Google

One day after Google DeepMind rolled out Genie 4 in Project Genie for Ultra subscribers, Google opened models/genie-4 in preview on 19 November inside the v1beta Gemini Interactions API and AI Studio, billing 240 video output tokens per frame (roughly $0.10 per second with audio) across 120-second sessions.

Paired on 19 and 20 November with zero-fee Tool Retrieval in the Interactions API, Gemini 4 Hydrogen robotics tools in Antigravity and NVIDIA’s open-source MotionBricks, Genie 4 moves interactive spatial simulation directly into standard SDKs.

Microsoft Ignite ’26 launch card set over a warm amber and slate-blue diagonal gradient
Image: Microsoft

At Ignite 2026 on 17 November, Microsoft introduced MAI Code 2, MAI Code 2 Flash, MAI Cyber 2 Flash ($0.10 input and $0.50 output, scoring 92.3% on CyberGym-E2E), MAI Image 2.7, MAI Realtime 1, Dragon Copilot Transcribe and microsoft-decision-1.1 alongside private Azure MAI Licensed Weights and Maia 300 accelerators (+40% tokens per watt).

Microsoft extended that stack on 19 November with Canvas on MAI Image 2.7 Flash while shipping GitHub Copilot security extensions and verified MCP connectors alongside third-party GPT-6.2 Astra support on Foundry.

05OtherAnthropicConfirmed

Anthropic prices 980M-share Nasdaq IPO at $83 under ticker ANT

Wed 18 Nov · Daily edition · 1 min read

Brushed-bronze Anthropic A-and-slash monogram on a dark marble pedestal in front of an ascending arc of illuminated emerald and amber glass market candlestick columns
Image: AI Compass illustration (Anthropic marks)

Anthropic priced its Nasdaq Global Select listing (ticker ANT) at $83 per Class A share on 18 November, topping its $70 to $78 roadshow band for an $81.34 billion base offering across 956.35 million primary shares and 23.65 million stockholder shares, plus a 147-million-share underwriter option.

Alongside its 19 November market debut, Anthropic kept shipping developer infrastructure across 16 to 17 November with experimental JavaScript Code Mode in Claude Code, iPhone Duo simulator testing and keyless Linked billing on the Claude Platform.

Glowing violet-and-white Qwen geometric emblem mounted on a dark obsidian MoE processor core routing sparse optical pathways to active GPU modules
Image: AI Compass illustration (Qwen marks)

Alibaba’s Qwen opened the 16 to 20 November slate with Qwen4-Max on QwenCloud, Qwen Studio, QwenWork and Qoder, combining 2.4 trillion backbone parameters, 71B active per token and 328B N-gram Embedding weights across a 1M multimodal window on OpenAI- and Anthropic-compatible endpoints.

Priced at half of Google Gemini 4 Neon’s $3 input rate and less than a fifth of GPT-6.2 Astra’s $8 tier, Qwen4-Max anchored a major Chinese MoE push alongside Tencent’s $0.070 Hy4-flash API and open-weight 770B Hy4.

07Open sourceThinking Machines LabConfirmed

Thinking Machines Lab open-sources 26B Inkling 2 Mini looped model

Fri 20 Nov · Daily edition · 2 min read

Video: Thinking Machines Lab

Released on 20 November on Hugging Face and Tinker under Apache 2.0, Inkling 2 Mini packs 26.4B total parameters (3.4B unique active, 5.4B looped active across 48 effective steps) into a roughly 20GB mixed NVFP4 and MXFP8 checkpoint running locally via llama.cpp, Ollama, Unsloth and Dynamo.

The release capped a strong 17 to 20 November open-weight run alongside Tencent’s open-weight 770B Hy4 and 220B Hy4-flash, Hugging Face’s Qwen3.8-27B-pi-adapted LoRA and MachinePulse’s Apache 2.0 Karpo Decision 31B.

Numbers and benchmarks of the week

Key specs, pricing and evaluations
ReleaseCompanyCategoryKey spec / benchmark / priceStory
Gemini 4 NeonGoogleModel1M unified window, #1 AA Coding Agent Index (58.6), $3 in / $15 out / $0.15 cache per 1M
GPT-6.2 Astra and SolOpenAIModelTied #1 Epoch Capabilities (174), Astra $8 in / $40 out / $0.40 cache (-20%), Sol $2 / $10
Genie 4 (Gemini API)GoogleModel720p 24 fps steerable WebRTC worlds + 48 kHz stereo audio, 120s cap, ~$0.10/sec
Qwen4-MaxQwenModel2.4T / 71B active MoE + 328B N-gram, 1M context, $1.5 in / $4.5 out / $0.075 cache per 1M
Microsoft MAI and Maia 300MicrosoftModel7 MAI models, MAI Cyber 2 Flash 92.3% CyberGym at $0.10/$0.50, Maia 300 +40% tokens/watt
Tencent Hy4 WeightsTencentOpen source770B / 49B active multimodal MoE + 3.3B DFly2 drafter, 1M context, open Hy License + FP8
Inkling 2 MiniThinking Machines LabOpen source26.4B / 5.4B looped MoE (~20GB NVFP4/MXFP8), 1M context, 48.2% DeepSWE v1.1, Apache 2.0
GPT-Live-1.1OpenAIModel128k full-duplex WebRTC and WebSocket voice API, $0.05/min active and $0.02/min standby
Liquid d1.2 and d1.2-ultraLiquid AIModel64k text, image and audio decision API, #1 AA Decision v2.1 (75), $0.02 and $0.042/1M input
NVIDIA MotionBricksNVIDIAOpen source350k+ mocap skills, 15,000 FPS at 2 ms on TensorRT for SOMA and Unitree G1, Apache 2.0 code

Category radar

Every other confirmed launch from 16 to 20 Nov

World models, speech, creative and decision APIs

Model

Coding agents, IDEs and software security

Coding agent

Open-weight MoEs, robotics and RL harnesses

Open source

Agent routing, web extraction and cloud runtimes

Infra

Computer use, enterprise plugins and workspaces

Product

Benchmarks, science, defences and legal oversight

Research

Weekend releases

Confirmed launches after Saturday morning

ResearchLean FRO and Mathlib InitiativeConfirmed

Lean and Mathlib certify 17M-line formal proof of the Riemann Hypothesis

The Lean FRO and Mathlib Initiative endorsed a joint community statement on 21 November certifying Levent Alpöge’s 17-million-line Lean proof of Mathlib’s RiemannHypothesis across Lean’s kernel, con-leche, con-ron and independent proof checkers.

SourceLean community blog: Lean verification of the Riemann Hypothesis · leanprover-community.github.io

InfraOpenRouterConfirmed

OpenRouter off-boards three inference providers and pulls 125 ZDR endpoints

Following canary-token data-handling findings by Dragos security researchers with no breach of OpenRouter itself, OpenRouter off-boarded three third-party inference providers, removed 125 endpoints from Zero Data Retention eligibility and notified affected users.

SourceOpenRouter: Data Processing Incidents on Inference Providers · openrouter.ai · More from OpenRouter

Coding agentGoogleConfirmed

Google resets Antigravity Gemini quotas alongside 22% caching gains

Google DeepMind’s Antigravity team reset Gemini usage quotas across all Antigravity accounts on 21 November and shipped prompt-caching improvements that fit 22% more engineering work inside each Gemini 4 Neon quota window.

SourceVarun Mohan (@_mohansolo) announcement on X · x.com · More from Google

ResearchArenaConfirmed

Arena places GPT-6.2 Astra sixth overall and third in coding on Agent Arena

Arena reported on 21 November that OpenAI’s GPT-6.2 Astra (Max) debuted at #6 overall on the Agent Arena Index, climbing five places to #3 on Coding (1,338 vs 1,288 for GPT-6 Astra) and +13 on Work (1,280) while dipping to #11 on Chat (1,211).

SourceArena (@arena) Agent Arena Index breakdown on X · x.com · More from Arena

That’s 16 to 20 Nov

84 launches from 16 to 20 November 2026, distilled and edited by .

Spotted an error? Email aiforsuccess.in+compass@gmail.com or tag @ai_for_success on X

Browse 16 to 20 Nov day by day

Full archive
  1. Monday 16 November

    Qwen launches 2.4T Qwen4-Max MoE model at $1.5 and $4.5 per million

    15 launches

  2. Tuesday 17 November

    Microsoft unveils seven MAI models, Maia 300, Microsoft IQ and Autopilot

    18 launches

  3. Wednesday 18 November

    Google launches Gemini 4 Neon at $3 and $15 per million tokens

    18 launches

  4. Thursday 19 November

    Google opens Genie 4 interactive world model in Gemini API

    16 launches

  5. Friday 20 November

    Thinking Machines Lab open-sources 26B Inkling 2 Mini looped model

    13 launches

Email Edition · Early Access

Get the 5-minute AI briefing

Daily and Sunday Weekly inbox editions open soon. Join the early-access list.

Zero spam · Privacy

! Please enter a valid email address (for example, you@domain.com).

Morning Edition Alerts

Get a 1-click alert at 11:00 IST

One quiet browser alert when each daily or Sunday Weekly edition goes live between 11:00 and 12:00 IST. No email required. Privacy