AI Compass

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

Tuesday 17 November 2026 · 18 launches · the whole day in 16 minutes

Microsoft unveils MAI models and Maia 300 as OpenAI ships GPT-6.2 Astra

Lead storyModelMicrosoftConfirmed

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

Spanning GitHub Copilot, Microsoft Foundry and private Azure MAI Licensed Weights alongside generally available Cobalt 200 VMs and Foundry IQ retrieval at Ignite on 17 November 2026.

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

Microsoft opened Ignite 2026 in San Francisco on 17 November 2026 with seven task-specialised MAI models, private Azure weight licensing, next-generation custom silicon, the Microsoft IQ context platform and autonomous Copilot modes. Across coding and security, MAI Code 2 and MAI Code 2 Flash reached general availability in GitHub Copilot, hitting 98% of GPT-6.1 Sol’s coding score at 40% of the cost and 96% of Claude Opus 5.5 at 5% of the cost, while MAI Cyber 2 Flash scored 92.3% on CyberGym-E2E at $0.10 input and $0.50 output per million tokens.

Multimodal MAI endpoints, Maia 300 silicon and Microsoft Retriever 1

MAI Image 2.7 ranked first on Artificial Analysis Image Editing v2.1 at $34.82 per 1,000 images alongside MAI Image 2.7 Flash ($10.24 per 1,000) and an interactive Canvas workspace. MAI Realtime 1 launched for speech-to-speech agents at $0.01 per audio minute, Dragon Copilot Transcribe built on MAI Transcribe 2 scaled across 10 million monthly clinical encounters, and multimodal microsoft-decision-1.1 hit general availability on Foundry, OpenRouter and Vercel.

Keynote slide displaying four custom Azure silicon packages side by side for Maia 300, Boost DPU, Integrated HSM and generally available Cobalt 200
Microsoft unveiled Maia 300 alongside Boost DPU, Integrated HSM and generally available ARM-based Azure Cobalt 200 VMs.Image: Microsoft

Under MAI Licensed Weights, enterprises fine-tune and host MAI checkpoints inside private Azure tenants. Chief executive Satya Nadella unveiled Maia 300 accelerators targeting 40% more tokens per watt than rival GPU clusters, confirmed Maia 200 in production at 1.7 times more tokens per watt, and shipped Cobalt 200 VMs and HorizonDB generally. Microsoft also launched Microsoft IQ with preview microsoft-retriever-1 cutting retrieval latency 85% to 0.76s, 24-hour Copilot Autopilot, Copilot Workspaces, Salesforce plugins and generally available Microsoft Perception security agents.

Architecture flow showing Foundry IQ Auto routing queries through microsoft-retriever-1 plan-and-search loops before escalating the hardest 20% to an Azure OpenAI LLM
Foundry IQ Auto offloads 80% of multi-turn search planning to microsoft-retriever-1, cutting median latency from 5.09s to 0.76s.Image: Microsoft

Top launches

Stories 2 to 10 of 18

02ModelOpenAIConfirmed

OpenAI launches GPT-6.2 Astra and GPT-6.2 Sol at $8 and $2 input pricing

Live on 17 November 2026 across the OpenAI API, Codex, ChatGPT Work, GitHub Copilot, Microsoft Foundry, OpenRouter and Perplexity with $0.40 cached Astra reads.

2 min read

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 released flagship GPT-6.2 Astra and everyday agentic model GPT-6.2 Sol on 17 November 2026 across the OpenAI API, Codex and ChatGPT Work for Plus, Pro, Enterprise and Edu workspaces, with day-one availability on GitHub Copilot, Microsoft Foundry, OpenRouter (including Pro tiers) and Perplexity Search and Computer. In the API, GPT-6.2 Astra drops 20% below GPT-6 Astra to $8 per million input tokens, $40 per million output tokens and $0.40 per million cached input reads (down 60% from $1), while GPT-6.2 Sol retains $2 input and $10 output per million tokens and expands ultrafast serving to Pro 200 plans.

Benchmark highs, token efficiency and coding-agent trade-offs

OpenAI reports that GPT-6.2 Astra sets new highs on Agents’ Last Exam (5.2 points above GPT-6 Astra at under one-third the task cost, and 5.5 points ahead of Claude Fable 5.2), Terminal-Bench-Science 0.1 (up 9 percentage points), AA-AnalystAgent spreadsheet tasks (62.5% accuracy across five runs), Harvey LAB-AA v1.1 legal deliverables and internal GDPval v2 workflows (up 6.2 percentage points using 22% fewer output tokens), extending OpenAI’s healthcare computer-use push into CAD and Blender physics simulation.

Epoch AI scores GPT-6.2 Astra at 174 on the Epoch Capabilities Index (tied with Claude Fable 5.2), and Artificial Analysis ranks Astra second on its Intelligence Index and first on AA-3D (1318) and AA-Robotics v0.1 (79). GPT-6.2 Sol finishes within 0.5 points of Astra on Cognition’s FrontierCode 2.0 at one-fifth the cost, though Artificial Analysis measured flat Coding Agent Index scores for Astra and a 3-point drop for Sol driven by a 14-point SWE-Atlas-QnA dip.

03ProductGoogleConfirmed

Google makes Gemini Tasks free in six countries on Gemini 4 Hydrogen

Rolling out free of charge on 17 November 2026 for eligible Gemini app users in the US, Japan, South Korea, Canada, Australia and India with AP2 Google Pay checkout.

1 min read

Illuminated four-pointed rainbow Gemini star emblem over dark obsidian glass dispatching cyan beams to wordless mail, calendar, phone and contactless card tiles
Image: AI Compass illustration (Google marks)

Google expanded Tasks inside the Gemini app free of charge on 17 November 2026 to hundreds of millions of eligible accounts starting in the United States, Japan, South Korea, Canada, Australia and India. Powered by Gemini 4 Hydrogen, Tasks separates multi-step administrative projects and recurring background schedules from quick conversational Chat threads.

Email-to-calendar RSVPs, local business calls and AP2 Google Pay checkout

Chief executive Sundar Pichai detailed four action capabilities included in the free-tier rollout. Users can forward messages directly to gemini@google.com so Gemini processes event RSVPs and populates a shared family Google Calendar, or instruct Gemini to telephone local businesses to compare price quotes and check item stock, with mandatory user approval before dialling and full call transcripts afterwards.

For web transactions, Tasks signs into websites through credentials stored in Google Password Manager and completes purchases securely through Google Pay built on the Agent Payments Protocol (AP2). Pichai noted that Google is partnering directly with major websites and security vendors so authenticated agent sessions operate smoothly for site owners alongside users. Google has not announced rollout dates for European Union or UK accounts.

04InfraPerplexityConfirmed

Perplexity launches Orchestrator API to route across 45+ AI models

Live on 17 November 2026 with OpenAI Responses, Chat Completions and Anthropic Messages compatibility, consolidated billing and perplexity/auto failover.

1 min read

Glowing white-and-teal Perplexity geometric crystal emblem mounted in a dark metal switchboard routing fiber-optic beams into six parallel processor modules
Image: AI Compass illustration (Perplexity marks)

Perplexity launched the Perplexity Orchestrator API on 17 November 2026, giving developers unified access to more than 45 open-weight and proprietary models from OpenAI, Anthropic, Google, xAI and open-source labs through one API key and consolidated billing. Following its 16 November Perplexity Monitor API launch, the gateway accepts OpenAI Responses, OpenAI Streaming Chat Completions and Anthropic Messages schemas and can attach Perplexity web search and inline citations in one call.

Task-aware perplexity/auto model selection and provider failover

Developers can pin an explicit model ID such as claude-fable-5-2 or target the perplexity/auto router, which inspects each request to balance task capability against cost and latency before selecting an underlying deployment based on live provider availability and capacity. If a primary deployment fails or throttles, Orchestrator automatically retries across healthy endpoints without client-side routing code.

Perplexity published per-model context windows, output token limits, prompt-caching rules, safety identifiers, Analytics API telemetry, AWS Marketplace billing options and Sonar migration guides across docs.perplexity.ai.

05Coding agentGoogleConfirmed

Google AI Studio adds custom domains and HD Android emulator builds

Shipped in AI Studio Build on 17 November 2026 with in-studio domain purchasing, Cloud Firestore state sync, user authentication and US Google Play publishing.

1 min read

Colourful S-ribbon Google AI Studio emblem on dark obsidian connecting a frameless browser window with a glowing globe icon to an upright smartphone showing database sync rings
Image: AI Compass illustration (Google marks)

Google rolled out web and mobile deployment upgrades to AI Studio Build (ai.studio/build) on 17 November 2026, adding direct custom-domain purchasing and hosting alongside expanded Android app creation tools. Developers building web apps inside Google AI Studio can now buy a domain name in place and publish generated sites onto a branded custom domain without exporting repositories to an external registrar or hosting dashboard.

HD audio Android emulator, Cloud Firestore sync and Play library publishing

For mobile builders, Google upgraded the in-browser Android app builder first introduced at Google I/O in May 2026, pairing cleanly with Google’s free zero-setup Maps Platform toolkit in AI Studio. The embedded Android emulator inside AI Studio now renders at high-definition resolution with live audio output so multimedia and voice interactions can be tested directly in the browser.

Generated Android applications can also provision Cloud Firestore databases and user authentication to persist and synchronise data across devices, wire Gemini calls securely without exposing client keys, and publish builds straight to a user’s personal Google Play library on Android phones in the United States. Google has not announced custom-domain pricing tiers or international Play library rollout dates.

06RoboticsGoogle DeepMindConfirmed

Google DeepMind launches Gemini 4 Hydrogen robotics plugin for Antigravity

Available on 17 November 2026 from Antigravity’s Build With Google shelf, combining USB robot bring-up, MuJoCo simulation and 84.8 ERQA spatial reasoning.

1 min read

Video: Google

Google DeepMind and Google Developers released the Robotics plugin for Google Antigravity on 17 November 2026, pairing gemini-4-hydrogen with built-in MuJoCo physics simulation to turn Antigravity into a conversational real-to-sim-to-real workbench. Installed from the Build With Google shelf under Customizations, the plugin bundles agent skills for USB driver installation, external depth-camera discovery, safe arm calibration, MuJoCo scene generation, grasp planning and closed-loop controller synthesis.

Ten-minute USB hardware bring-up and 84.8 ERQA embodied reasoning

Starting from an unconfigured desk, asking Gemini 4 Hydrogen in Antigravity to bring up a USB-connected I2RT YAM arm, a low-cost SO-ARM101 arm or an industrial dual-arm FR3 Duo installs dependencies, locates cameras, calibrates joints and verifies test motions in roughly 10 minutes before running plain-language manipulation prompts.

Following Gemini Robotics 2.5, Google folded Gemini Robotics ER 2 spatial capabilities directly into gemini-4-hydrogen, lifting average ERQA embodied reasoning from 75.5 to 84.8 at high thinking and scoring 58 on Artificial Analysis Robotics 0.1 across eight 6-axis tabletop tasks, matching Gemini 4 Argon at lower token cost.

Eleven-axis ERQA radar chart showing Gemini 4 Hydrogen averaging 84.8 versus 75.5 for Gemini Robotics ER 2 across gauge reading, semantic keypoints, error detection and safety
Gemini 4 Hydrogen averages 84.8 on the 11-category ERQA embodied reasoning benchmark, ahead of Gemini Robotics ER 2 at 75.5.Image: Google

07Coding agentCognitionConfirmed

Cognition launches Devin Bots to lead parallel teams of coding agents

Persistent, memory-backed engineering leads inside Devin break ongoing goals into parallel agent sessions, review pull requests and run post-deploy QA from 17 November 2026.

1 min read

Three hardware-styled Devin Bot visors in violet, cyan and coral glowing with dual LED eye matrices against a dark backdrop
Image: Cognition

Cognition introduced Devin Bots on 17 November 2026, adding persistent, goal-owning manager agents that coordinate multiple Devin coding instances in parallel. Building on Devin Apps and the upgraded Devin Teams plan, a Devin Bot retains cross-session memory, inspects repositories and production telemetry, decomposes long-running objectives into parallel Devin tasks and follows up on stalled work.

Autonomous PR reviews, post-deploy QA and building Devin’s Android app

Unlike single-ticket coding sessions that stop after opening one pull request, Devin Bots inherit workspace integrations and permissions out of the box to triage overnight incidents, review pull requests, run quality assurance after deployments to catch regressions, and deliver morning briefings listing only architectural or product choices that need human judgement.

Cognition reported that internal Devin Bots merged roughly 1,600 pull requests in one week across ongoing responsibilities such as open-source maintenance, release management and inference optimisation, and autonomously built Devin’s Android app after an engineer assigned an Android Bot the project on 28 September 2026 with zero dedicated Android staff. Cognition has not announced standalone Devin Bots pricing.

08Coding agentAnthropicConfirmed

Anthropic adds experimental Code Mode to Claude Code for JS tool chains

Available on 17 November 2026 in Claude Code v2.1.323+ for direct Claude and Anthropic Console accounts via the CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1 flag.

1 min read

Warm terracotta Claude sunburst emblem inside a dark-glass terminal chassis routing amber and cyan conduits through search, file-stack and three parallel gear cubes
Image: AI Compass illustration (Anthropic marks)

Anthropic launched experimental Code Mode in Claude Code v2.1.323 and later on 17 November 2026, opted into via the environment variable CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1 in the terminal. Rather than issuing one tool invocation per model round-trip, Claude writes a JavaScript program that chains multiple tool calls in one step, loops across files for repetitive edits, filters bulky search results before reading them and runs independent commands concurrently.

Background execution, 10-minute timeouts and direct API constraints

While background commands run, Claude Code can continue answering user prompts in the active terminal session, extending recent Claude Code workflow updates. Foreground scripts stop after a 10-minute timeout, Auto Mode’s safety classifier inspects the generated JavaScript program rather than individual calls, and PreToolUse or UserPromptSubmit hooks can refuse calls or withhold script output.

Code Mode requires a direct Claude subscription or Anthropic Console account on either release channel and cannot be resumed outside the terminal, moved to Desktop, forked into background sessions or used with Remote Control, Claude in Chrome, computer use, Bedrock, Vertex or Microsoft Foundry.

09ProductOpenAIConfirmed

OpenAI introduces ChatGPT for Sales with 30+ CRM and revenue plugins

Launched on 17 November 2026 with a first-party OpenAI Sales plugin, persistent dot delegation and connectors for Salesforce, HubSpot, Gong, Clay and Docusign.

1 min read

Glowing white hexagonal OpenAI blossom emblem on dark obsidian linked by gold and emerald conduits to funnel, voice-headset, slide-chart and contract-seal glass tiles
Image: AI Compass illustration (OpenAI marks)

OpenAI introduced ChatGPT for Sales on 17 November 2026, pairing a new OpenAI-built Sales plugin with more than 30 new and updated partner plugins so revenue teams can query CRM records, call transcripts and buyer signals in one workspace. The first-party Sales plugin supplies structured skills for account prioritisation, discovery call prep, proposal drafting and follow-ups, while persistent dot agents can run those plugins between chats to surface deal changes.

Six workflow categories from CRM and prospecting to CPQ contracting

Partner connectors span Salesforce, HubSpot, Close CRM, monday CRM, Pipedrive, Zoho CRM and Attio for pipeline context; Apollo, Clay, Common Room, ZoomInfo, Rox, Actively AI, Outreach and Salesloft for buyer signals and sequences; and Gong, Fireflies, Otter.ai, Granola, Zoom and the ChatGPT Meetings plugin for call notes.

Six-card OpenAI grid grouping built-in ChatGPT tools and 30+ partner logos across content enablement, CRM accounts, prospecting, conversations, deal risk and quotes
ChatGPT for Sales organises built-in OpenAI tools and 30+ partner plugins across six end-to-end revenue workflows.Image: OpenAI

Teams can analyse deal risk via Backstory, Hex, Tableau, Power BI, Zoom Revenue Intelligence and the ChatGPT Data plugin, rehearse objections in ChatGPT Voice, generate decks or ROI calculators via ChatGPT Slides, Sites, Gamma, Figma and Canva, and draft quotes in Docusign, Ironclad, PandaDoc and ServiceNow CPQ. OpenAI has not announced separate per-seat pricing for ChatGPT for Sales.

10Open sourceHugging FaceConfirmed

Hugging Face releases Qwen3.8-27B-pi-adapted RL LoRA harness adapter

Open-weight TRL adapter from harness-hub on the Hugging Face Hub uses safeagents rule checks to lift SWE-bench Pro Hard 51 pass@1 from 29.4% to 62.7% on 17 November 2026.

1 min read

Hugging Face Harness Adapters graphic showing concentric orbital rings for Hermes, opencode, pi, LFM2.5-24B, LFM3 and TwIL-8B next to the Hugging Face mascot
Image: Hugging Face

Hugging Face’s harness-hub team released open-weight Qwen3.8-27B-pi-adapted and its safe reinforcement learning methodology on 17 November 2026, training lightweight per-harness LoRA adapters directly inside unmodified local coding harnesses including pi, opencode and hermes. Building on earlier LFM2.5-24B-A2B, LFM3-1.7B and TwIL-8B runs alongside Hub OpenRouter inference routing, one shared base server can host adapters for multiple agent harnesses simultaneously.

Repo2RLEnvs trajectory tasks, safeagents proxy and pi coding benchmarks

Training mixes SmolDataEnvs with Repo2RLEnvs, which converts open-source coding sessions shared on the Hugging Face Hub into verifiable build-heavy RL environments. Every tool call passes through safeagents, a deterministic rule-based proxy running in 10 to 20 ms that blocks reward-hacking shortcuts and assigns violating rollouts a zero score.

Served over the openai-completions endpoint via vLLM, Transformers or llama.cpp, Qwen3.8-27B-pi-adapted raised pass@1 on SWE-bench Pro (Hard 51) from 29.4% to 62.7% and on Terminal-Bench 2.1 from 39.5% to 55.8%, while cutting tool calls per task by 55% to 62%.

Also that day

Shorter items

11ResearchIsomorphic LabsConfirmed

Isomorphic Labs raises $5B Series C to scale its AI drug design engine

Alphabet spin-out Isomorphic Labs and founder Demis Hassabis announced on 17 November 2026 that the London AI drug discovery company has raised $5 billion in Series C funding co-led by Greenoaks and Sequoia Capital alongside Andreessen Horowitz, Alphabet and all existing investors to scale its Isomorphic Labs Drug Design Engine (IsoDDE) and advance internal therapeutic candidates towards clinical trials.

SourcesIsomorphic Labs announces Series C investment round · Demis Hassabis (@demishassabis) $5B funding announcement on X

12OtherNVIDIAConfirmed

NVIDIA founder Jensen Huang says he would return as CEO if board asks

Five days after stepping down on 12 November 2026 when CFO Colette Kress became interim chief executive, NVIDIA founder Jensen Huang published a statement on X on 17 November owning serious mistakes in his personal life, pledging full cooperation with the Board of Directors’ independent investigation, and stating he would be honoured to return as CEO if the board decides it serves shareholders, employees, NVIDIA and the United States.

SourceJensen Huang (@JensenHuang) personal statement on X

13Coding agentGitHubConfirmed

GitHub previews Copilot Chrome extension and Azure DevOps app support

GitHub announced at Microsoft Ignite on 17 November 2026 that the GitHub Copilot app now supports Azure DevOps repositories in public preview (including branch creation, review-comment fixes and automated CI agent merge), a public-preview GitHub Copilot browser extension for Chrome and Edge running inside an isolated tab group, and generally available GitHub Enterprise Server integration.

SourcesGitHub Copilot app, Microsoft Ignite releases - GitHub Changelog · GitHub (@github) Copilot Chrome extension preview post on X · GitHub (@github) Copilot Azure DevOps integration post on X

14InfraAnthropicConfirmed

Anthropic opens Linked billing beta on the Claude Platform without keys

Anthropic launched Linked billing in beta on the Claude Platform on 17 November 2026 via the user-profiles-billing-2026-10-01 beta header, enabling users of third-party tools such as Factory, Linear and Ramp to link their direct-billed Claude Platform organisation, set spend caps or unlink apps in the Claude Console, and pay for Claude API calls without sharing raw API keys.

SourcesClaude Platform release notes (17 November 2026) · Claude Developers (@ClaudeDevs) Linked billing beta post on X

15ProductPerplexityConfirmed

Perplexity rolls out Multiagent mode in Computer and Teammates for Max

Perplexity rolled out a Multiagent effort level on the web on 17 November 2026 for Perplexity Max subscribers using Computer and Teammates, allowing a lead orchestrator model such as Claude Fable 5.2 to split complex workflows across parallel subagents running models like Gemini 4 Carbon, GPT-6.1 Sol and Claude Opus 5.5 while sharing intermediate findings into one synthesised answer.

SourcesPerplexity (@perplexity_ai) Multiagent rollout post on X · Aravind Srinivas (@AravSrinivas) Multiagent Max details on X

16Open sourceMachinePulseConfirmed

MachinePulse open-sources Karpo Decision 31B single-pass decision models

MachinePulse released the six-model open-weight Karpo Decision family (0.8B to 31B, built on Gemma 4 31B under Apache 2.0; research blog published 13 November 2026 and announced on X on 16 November), using branch attention and a direct readout head to answer multiple typed questions over text and images in one forward pass, scoring 73.9 on Decision Index Full at 95 ms median latency on one RTX PRO 6000.

SourcesKarpo Research: Karpo Decision 31B: A Fast, Accurate Multimodal Decision Model · MachinePulseAI/karpo-decision-31B on Hugging Face

17ModelElevenLabsConfirmed

ElevenLabs releases Scribe v3 Live Medical for real-time clinical calls

ElevenLabs launched Scribe v3 Live Medical inside ElevenAgents on 17 November 2026, adapting its 150-millisecond real-time Scribe v3 Live speech-to-text model for live patient, member and provider phone calls with custom vocabularies of up to 10,000 medication, procedure and clinician names and the lowest word error rate among medical models it tested on PriMock57.

SourcesElevenLabs: Scribe v3 Live Medical: transcription for healthcare agents · ElevenLabs (@ElevenLabs) Scribe v3 Live Medical post on X

18ResearchGitHubConfirmed

GitHub and Microsoft Security triage 9,791 AI bug scans in Codename MDASH

GitHub and Microsoft Security detailed the Codename MDASH agentic code-security pilot at Microsoft Ignite on 17 November 2026, filtering 9,791 raw AI vulnerability findings across six open-source repositories into 99 human-verified private advisories complete with suggested patches so open-source maintainers can ship fixes without triaging unvetted AI bug-report spam.

SourcesGitHub Blog: Maintainers don’t need more AI bug reports. They need fixes. · GitHub (@github) Codename MDASH report post on X

Edition Alerts

Never miss the morning briefing

Get one quiet browser alert when the new edition goes live between 11:00 and 12:00 IST. No email required. Privacy

That’s the day

18 launches from Tuesday 17 November 2026, edited by .

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

Recent days

The archive
  1. Wednesday 18 November

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

    18 launches

  2. Monday 16 November

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

    15 launches

  3. Friday 13 November

    OpenAI launches ChatGPT Wallet for tokenised agent payments in dots

    15 launches

  4. Thursday 12 November

    Google launches Gemini 4 Hydrogen for everyday coding and agents

    13 launches

  5. Wednesday 11 November

    Google launches Gemini Decisions API and Gemini Decision 1 at $0.10 per million input tokens

    13 launches