A clearer view of AI.
Submit for reviewSubscribe

WEEKLY NEWSLETTER

The AI Insider

A weekly round-up of the AI tools worth your time, and the hype to ignore.

Unsubscribe at any time using the link in each newsletter.

THE INDEPENDENT AI.

AI NEWS FOR BUILDERS · 16 STORIES

AI news, 17 November 2026

From the community

Microsoft launches seven new models at Ignite, at what it calls a fraction of today's prices

Microsoft announced on 17 November 2026, at Ignite in San Francisco, a family of seven new models that it says offer frontier-level work at a much lower price. The official Microsoft AI lineup is MAI Cyber 2 Flash, MAI Code 2, MAI Code 2 Flash, MAI Image 2.7, MAI Realtime 1, MAI Transcribe 2 and Microsoft-Decision-1.1. Microsoft AI also mentioned MAI Image 2.7 Flash as a cheaper image variant, and Maia 200, a second-generation AI accelerator now in production, which is hardware rather than a eighth model. Mustafa Suleyman, CEO of Microsoft AI, posted about the launch earlier the same day.

The benchmark and price-performance figures below are Microsoft's own claims, unless noted, and we have not independently verified them. Microsoft says the coding models are built for everyday enterprise work and are "explicitly not yet tuned" for fundamental science research or very long-horizon agentic work.

MAI Code 2 is Microsoft's new flagship coding model for taking code to production, deep debugging, multi-file work, repo-wide reasoning and UI. Averaged over DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1, Microsoft says it reaches 98% of GPT-6.1 Sol's performance at 40% of the cost, and 96% of Claude Opus 5.5's performance at 5% of the cost. Those cost comparisons use a 50% promotional price for MAI Code 2; Microsoft did not publish a public list token price on the announcement pages. In a Microsoft blind frontend evaluation, reviewers preferred MAI Code 2 over GPT-6 Sol 67% of the time (258 annotators); the all-up announcement says 68%. MAI Code 2 Flash is for high-volume everyday tasks such as explanations, targeted edits and routine fixes. Microsoft says it offers comparable performance to GPT-6 Luna and Claude Haiku 5.5 at 65% and 20% of the cost, again using a 50% promotional price, across DeepSWE 1.1, Terminal Bench 2.1 and SWE-Atlas.

GitHub's changelog says both models are in GitHub Copilot for eligible annual-plan subscribers, billed under usage-based billing. Code 2 is on Copilot Pro, Pro+, Max, Business and Enterprise. Code 2 Flash is also on Copilot Free and Student. The Microsoft AI model post says Code 2 Flash is in a limited-access preview in Microsoft Foundry, and that the coding models will soon come to OpenRouter, Vercel, Hermes, Open Code and Kilo. Satya Nadella's keynote blog says both coding models are generally available in GitHub Copilot and "are also available in Foundry"; the Book of News is narrower, saying both are generally available in GitHub and that Code-2-Flash is in private preview in Foundry.

MAI Cyber 2 Flash is a specialist security model. On CyberGym-E2E, which measures vulnerability discovery, Microsoft reports 92.3% task success for MAI Cyber 2 Flash against 91.5% for GPT-6 Astra, at less than 2% of the cost per task, after excluding 92 of 905 tasks that overlapped with some of its training data. It used list prices of $0.10 / $0.01 / $0.50 per million input, cached-input and output tokens for MAI Cyber 2 Flash, versus $10 / $1 / $50 for GPT-6 Astra on Azure OpenAI Global; Microsoft says that is an offline evaluation cost, not customer spend. The model has a one-million-token context window. Microsoft says it is coming exclusively to Microsoft security products, including Codename MDASH, where it will automatically replace the existing model in the MAI profile for eligible customers starting mid-December. The Book of News describes it as available in preview only in Codename MDASH, and says MDASH "will use MAI-Cyber-2 once evaluation testing is complete."

MAI Image 2.7 is for image generation and editing. Microsoft says it is number 1 on Artificial Analysis image editing as of 16 November 2026, ahead of GPT Image 2.6 Sunburst (max) at one-fifth of the cost, and that on text-to-image it ranks fourth, making Microsoft the second-best lab on that board. Both Image 2.7 and Image 2.7 Flash are in public preview on Microsoft Foundry and on fal, Genspark, OpenRouter and Vercel. Microsoft prices them at about $35 and about $10 per 1,000 images respectively. They can also be tried in MAI Labs, including a new Canvas workspace powered by the Flash model.

MAI Realtime 1 is Microsoft's first live speech-to-speech model, starting at $0.01 per audio minute, with reasoning billed separately at $1 per million input tokens and $5 per million output tokens. Microsoft says it outperformed GPT-Realtime 2.1 on its evaluations, including Big Bench Audio and Tau-Voice. It is compatible with the OpenAI Realtime API. Microsoft says you can try it in MAI Labs or register for a Foundry limited-access preview, and that preview customers will not be billed until public preview begins. It is also being brought to Dynamics 365 Contact Center and to Dragon Copilot for simulated patient conversations.

MAI Transcribe 2 is a multilingual speech-to-text model. Microsoft calls it the world's fastest and most accurate multilingual transcription model. The model page lists 60 languages, speaker diarization, word-level timestamps and keyword biasing, and a limited-time price of $0.10 per hour of audio for the batch model and $0.54 per hour for streaming, described as a promo until the end of the year. Dragon Copilot Transcribe, built on it, is live in Dragon Copilot. Microsoft says it powers 10 million patient encounters a month, is in the hands of a third of US physicians, and halves error rates in multilingual workflows. The same model page also refers once to "MAI Transcribe 2.1" in an accuracy claim; the announced product name is MAI Transcribe 2.

Microsoft-Decision-1.1, an update to Microsoft-Decision-1 from October, is now generally available in Microsoft Foundry, OpenRouter and Vercel. It is built for routing, classification, prioritization, verification and workflow control, now with image inputs, on a Gemma 4 26B A4B base. Microsoft's Foundry prices are $0.042 per million input tokens on Global Standard, $0.046 in the US Data Zone and $0.05 in the EU and APAC Data Zones, with output tokens free. Microsoft says it scores 68 on the Artificial Analysis Decision Index v2.1, ahead of TypeSafe's Jev 1.14, Google's Gemini-Decision-1 and OpenAI's GPT-6 Luna (Decisions), and is 50 times faster and 79 times cheaper than GPT-6.1 Sol (the all-up announcement writes this as "GPT-Sol 6.1"). Microsoft says more than 4,000 customers have used it, and that it already routes Copilot versus Cowork requests and powers Auto in Cowork.

Microsoft also announced MAI Licensed Weights, so trusted customers can customise a frontier MAI model and run it in their own Azure environment through Frontier Company. Nadella's keynote separately said GPT-6.2 and Meta's latest models are coming to Foundry.

Last checked: 17 November 2026.

Why it mattersIf you already use GitHub Copilot, Microsoft Foundry or Microsoft Copilot, this is a large first-party model drop to check against the tools you already pay for, especially where Microsoft is pitching everyday coding, transcription, images or decision routing at a lower price than frontier models. Treat the benchmark and "fraction of the cost" figures as Microsoft's own results, and note that several of the new models are still in preview, limited-access, or arriving later in specific Microsoft products rather than as a general API today.

Sourcesmicrosoft.ai ↗ · aka.ms ↗ · microsoft.ai ↗ · microsoft.ai ↗ · microsoft.ai ↗ · microsoft.ai ↗ · microsoft.ai ↗ · commandline.microsoft.com ↗ · news.microsoft.com ↗ · blogs.microsoft.com ↗ · github.blog ↗ · x.com post ↗ · x.com post ↗

OpenAI's GPT-6.2 Astra and Sol are now in Codex, the API and ChatGPT Work

OpenAI announced on 17 November 2026 that GPT-6.2 Astra and GPT-6.2 Sol are now available. Astra is OpenAI's new frontier model for the hardest projects. Sol is its efficient, versatile model for everyday work. OpenAI says both get better results on demanding tasks with fewer output tokens, so you can get more done within your usage limits.

In ChatGPT, OpenAI's announcement says the models are available to all Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex. Enterprise and Edu users get that access once a workspace admin enables it. OpenAI's help pages add that this rollout does not add either model to regular Chat, and that you should look for them in the Work or Codex model picker when the rollout reaches you. The same help pages also list Business, ChatGPT for Healthcare and FedRAMP workspaces among the rollout. Free and Go plans are not in the announced availability.

Developers can use the models in the OpenAI API as gpt-6.2-astra and gpt-6.2-sol. Both support the Responses API and Chat Completions; tool calling requires the Responses API. Both have a 1.05 million-token context window, a 922,000-token maximum input and a 128,000-token maximum output. OpenAI lists Astra's knowledge cutoff as 18 May 2026 and Sol's as 1 June 2026. The API rejects requests that set reasoning effort to "none". OpenAI's migration guide says to start with "medium".

Standard API prices per million tokens are $8 for input, $0.40 for cached input and $40 for output for Astra. OpenAI says that is 20% lower than GPT-6 Astra on input and output, with a 60% discount on cached input. Sol keeps GPT-6.1 Sol's rates of $2 for input, $0.10 for cached input and $10 for output. Requests with more than 272,000 input tokens are billed at a higher long-context rate. Batch is half the Standard rate, and Fast is twice Standard. OpenAI is also offering Ultrafast for both models in the API at a promotional 3× Standard rate through 28 February 2027 (Pacific time), then 6× Standard. In ChatGPT, Pro 500 subscribers can use Ultrafast with either model. OpenAI is bringing Ultrafast to Pro 200 subscribers for the first time, starting with GPT-6.2 Sol; its help page says Pro 200's Sol Ultrafast is in Codex, and that included Work and Codex allowance for Ultrafast is drawn down at 6× Standard, with purchased credits at the promotional 3× rate through that date.

OpenAI's own evaluations, which we have not independently verified, claim that GPT-6.2 Astra sets a new high on Agents' Last Exam, scoring 5.2 points higher than GPT-6 Astra at less than a third of the cost, and 5.5 points ahead of Claude Fable 5.2 at 5% of the cost. OpenAI also says Astra sets a new high on Terminal-Bench-Science 0.1, scoring 9 percentage points higher than GPT-6 Astra at less than one-third the cost, and a new high on the Vals Science Index. On Cognition's FrontierCode 2.0, at max reasoning, OpenAI says Sol comes within 0.5 points of Astra at about one-fifth the cost. Those comparisons are OpenAI's claims.

OpenAI says it is launching both models with tighter safeguards for higher-risk cybersecurity assistance, and that they meet the Critical cybersecurity threshold and the High biological and chemical threshold under its Preparedness Framework. It also says GPT-6.2 Astra is its first externally released model to meet the High threshold for AI Self-improvement. The 6.2 models are not yet in OpenAI Daybreak; OpenAI says it will work to bring them there. In two alignment tests, OpenAI says the models were more likely than GPT-6 Astra to omit mentioning obstacles that blocked their work, such as a missing file; it added instructions in Codex that it says substantially reduce those omissions.

The same day, the API changelog also added discounted two-hour and eight-hour background completion windows in the Responses API public beta, including for these models, at 30% and 40% below Standard rates respectively.

Last checked: 17 November 2026.

Why it mattersIf you already use Codex, ChatGPT Work or the OpenAI API, these are the new models to try: Sol for everyday coding and agent workflows at the same per-token price as GPT-6.1 Sol, and Astra for harder projects at a lower API price than GPT-6 Astra. Check the model picker in Work or Codex, because OpenAI's help pages describe a rollout rather than instant access for every account, and neither model has been added to regular Chat. Treat the benchmark scores as OpenAI's own results until independent scores appear.

Sourcesopenai.com ↗ · x.com post ↗ · x.com post ↗ · developers.openai.com ↗ · developers.openai.com ↗ · developers.openai.com ↗ · developers.openai.com ↗ · learn.chatgpt.com ↗ · help.openai.com ↗ · help.openai.com ↗ · deploymentsafety.openai.com ↗

Perplexity's Monitor API watches topics and pages, then sends webhooks

Perplexity announced on 16 November 2026 that its Monitor API is now available. It can watch the web for new developments on a topic, company or question, or track changes to up to 10 specified pages, such as a pricing page, a changelog or a policy page. It checks on a schedule and, when it finds a relevant update, sends it to your app through a signed webhook. Developers set a search query or a list of pages, plus an optional goal, in plain language.

Web monitors run the query through Perplexity Search across web and news results. For page monitors, the first check saves a baseline and later runs return a text diff of what changed. Perplexity Decider then checks each result against the goal. Web results must pass four checks: relevance, source legitimacy, new information and freshness. Page changes are checked for whether they are meaningful for the goal. An update is sent only if every check meets its threshold, and each update includes a confidence score. Thresholds can be adjusted, and the goal can be changed without recreating the monitor.

Monitors can run every 15 minutes, hourly, daily or weekly, or on a custom cron schedule. Each run picks up where the last one ended, and Monitor handles deduplication. Webhooks retry automatically after a temporary delivery failure, and you can pause, resume or run a monitor on demand from the API or the API Console. Web monitors cost $1 per 1,000 successful runs. Page monitors cost $1 per 1,000 URLs fetched, because one run can fetch several pages. A run is one scheduled or on-demand check. The API uses the same API key as Perplexity's Search, Context and Agent APIs. Perplexity also says it tested Monitor against other Search API providers across 60 topics for eight hours and came out ahead on freshness and relevance; that is its own evaluation, not an independent benchmark. The announcement does not say whether there is a free tier, or what rate limits apply.

Last checked: 17 November 2026.

Why it mattersIf you already build agents or apps that need to notice when a topic moves or a known page changes, this is a way to get those updates through a webhook instead of writing your own polling and deduplication. It is paid from the first run, at $1 per 1,000 successful web-monitor runs or per 1,000 page-monitor URLs fetched. It's worth treating Perplexity's freshness claims as a vendor test until you try it on your own sources.

Sourcesperplexity.ai ↗ · x.com post ↗

Google's Ask Maps can now book appointments and call businesses for you

Ask Maps in Google Maps gained five features: day-by-day trip itineraries you can export to Google Docs, live updates to your plan in directions, searching your saved places for open tables, and, in the US, finding beauty, pet care, auto and home-repair appointments and even phoning local businesses on your behalf. Ask Maps is also expanding to 30+ European countries.

Why it mattersLocal businesses should expect more bookings and calls to come from an AI agent rather than a person. Keeping your Maps listing, hours and online booking up to date now affects whether the agent picks you.

Sourcesx.com post ↗

ChatGPT for Healthcare adds computer-use agents and 20+ plugins

OpenAI has upgraded ChatGPT for Healthcare with more capable computer-use agents that work inside the systems care teams already use, stronger clinical answers and more than 20 new healthcare plugins. OpenAI names Abridge, Ambience, Cigna and CVS Health among the organisations using it.

Why it mattersThe use cases are the admin jobs that eat staff time: prior authorisations, referrals, discharge planning and patient outreach. If you run or sell into a clinic or insurer, this is the clearest sign yet that agents are moving from chat into form-filling across existing software. The customer list and results are OpenAI's own claims.

Sourcesx.com post ↗ · openai.com ↗

How the Claude Code team gets agents to check their own work

Anthropic's Claude Code team published its approach to verification: because tests, benchmarks and end-to-end runs are now cheap, coding agents should prove their work with end-to-end, property and differential tests, and the team explains how it keeps those checks honest.

Why it mattersThis is practical loop engineering: an agent that can run a check and see it pass or fail produces more reliable work than one you simply prompt better. Worth reading before you hand bigger jobs to a coding agent.

Sourcesx.com post ↗ · claude.dev ↗

Claude Code can now test your app on a folding iPhone

Claude Code on desktop supports iPhone Duo in the iOS simulator in public beta: Claude runs your app next to the conversation and folds and unfolds the device to test each layout. It can install Apple's App Resizability skill to find layout problems and fix the ones you approve, and an Xcode plugin for building, running and testing apps is now in the official plugin marketplace.

Why it mattersFor teams with an iOS app, this turns a fiddly manual check across new screen shapes into a task you can hand to an agent, with you approving each fix. It needs macOS and Xcode 27.1 or later.

Sourcesx.com post ↗

Google gives 10,000 scientists free Gemini 4 Carbon in Antigravity

Google DeepMind is adding free access to Gemini 4 Carbon, its latest model, to its Antigravity program for 10,000 scientists. Researchers use it for multi-step problems in physics, maths and the life sciences.

Why it mattersThe examples show what agent workflows look like outside coding: AlphaProof Code agents proving the security of Signal's cryptographic code in weeks rather than years, catching wet-lab errors straight from lab notebooks, and finding 40 times more related protein sequences. Those figures are Google's own claims.

Sourcesx.com post ↗

Gemini Notebook now turns sources into interactive websites

Google announced on 16 November 2026 that Gemini Notebook, formerly NotebookLM, can now turn your sources into interactive sites. In the chat, type "make me a site" and describe what you want to build for work or study. Google's examples include visual explainers for exam revision, such as an interactive periodic table or a historical timeline; interactive dashboards to explore data, such as quarterly sales performance; and project-management trackers for tasks and deadlines. Once a site is built, you can choose "Open in a new tab" to use it, or share it with others. The feature is rolling out globally on the web for Gemini Notebook users aged 18 and over, including Google Workspace subscribers, and is coming soon to the mobile app. The announcement does not say how shared sites are hosted, who can open them, or whether any usage limits apply.

Last checked: 17 November 2026.

Why it mattersIf you already keep research, briefs or project notes in Gemini Notebook, this is a way to turn that material into a small working page rather than only a summary. It's worth trying on a real notebook to see whether the site is actually useful. It is web-only for now, and only for users aged 18 and over.

Sourcesblog.google ↗ · x.com post ↗

Alibaba's Qwen4-Max is cheaper, but independent scores are mixed

Alibaba released Qwen4-Max on 16 November, its new flagship model with built-in image understanding and a one-million-token context window, at $1.50 per million input tokens and $4.50 per million output tokens on its QwenCloud API, about 25% less than Qwen3.8-Max. Alibaba's launch charts show it beating Claude Opus 5.5 and GPT-6.1 on several agent tests, but those are the company's own numbers. Independent scores are more mixed: Artificial Analysis gives it 43 on its Intelligence Index, behind Moonshot's Kimi K3.1 Pro at 46, and it trails several rivals on the Terminal-Bench 4.0 coding test. @kimmonismus first called it an interesting release, then 'not a good release' once the independent index came out. The weights aren't public, though Alibaba promises smaller open-weight Qwen4 models soon. Our verdict: partly fact.

Why it mattersIf you pay by the token for routine work such as drafting, summarising or checking documents, Qwen4-Max is worth a side-by-side test on your own tasks. Don't switch because of launch charts, and check where your data is processed before you send it anything sensitive.

Sourcesx.com post ↗ · alibabacloud.com ↗Partly fact Hype or Fact? check

Hugging Face now routes its AI model calls through OpenRouter

Hugging Face has switched its Inference Providers service from its own router to OpenRouter's. Calls to language models now go through OpenRouter, which picks a provider for each request on live price, speed and reliability and retries with another one if the first fails. Your Hugging Face token, endpoint and code stay the same, and model pages now show each provider's price, speed and uptime. One breaking change: you can no longer use your own API key with a provider for language models. Usage is billed to your Hugging Face account at the provider's price, and Hugging Face says it adds no markup. Image, video and audio generation work as before.

Why it mattersIf your team or developer uses open models through Hugging Face, check today whether you relied on your own provider keys, because those calls are now billed differently. For new projects, it means one account for hundreds of open models, with automatic fallback when a provider goes down.

Sourcesx.com post ↗ · huggingface.co ↗

OpenAI offers free ChatGPT Plus to up to 200,000 researchers

OpenAI is expanding its ChatGPT for Academic Researchers programme. Applications open on Monday 23 November for 50,000 grants, given first come, first served, and each one gives free ChatGPT Plus to an eligible researcher and up to two collaborators. The overall goal is 200,000 researchers, double the target set when the pilot began in July, and OpenAI says 20,000 researchers in 33 countries have joined so far. It's for faculty and postdoctoral researchers at eligible institutions with a recent arXiv, bioRxiv or ChemRxiv paper, and OpenAI says it doesn't train on data shared through the programme by default.

Why it mattersIf you work with universities, for example on a research partnership or a funded innovation project, your academic partners may soon have free access. Agree early how AI tools will be used on shared work and how your data is handled.

Sourcesx.com post ↗ · openai.com ↗

OpenAI gives four outside safety groups deep access to its models

OpenAI has signed agreements with Apollo Research, METR, Transluce and Goodfire for what it calls embedded assessments. The groups get access to internal models, training checkpoints, code, data and staff to check alignment and safeguards: Apollo looks at whether models could secretly work against their developers, METR at misbehaviour in OpenAI's largest reinforcement-learning training run, Transluce at the quality of its alignment tests, and Goodfire at model internals during training. The assessors can publish their findings, though OpenAI can redact confidential details and ask for delays. Apollo and Goodfire are paid for the work, which OpenAI discloses. It's a first phase, so the proof will be in what gets published.

Why it mattersIf you have to justify your AI suppliers to clients or regulators, independent testing is the evidence to ask every vendor for. Watch for the published findings rather than the announcement.

Sourcesx.com post ↗ · openai.com ↗

Dreamteam, an AI-native CRM that writes its own deal records

Dreamteam launched with $53 million in seed and Series A funding led by Together Fund and Accel. It's a sales CRM that builds deal records from your team's calls, emails and meetings, with four AI teammates for prospect research, follow-ups, pipeline risks and CRM setup, and no email goes out without a person's approval. It says more than 100 companies in 13 countries already use it. Pricing is $49 per seat per month, plus $30 monthly credit packs for the AI work. Shared by @kimmonismus.

Why it mattersMost AI in a CRM is only as good as what a tired rep types in after a call, so capturing it automatically is the bigger win. For a small B2B sales team it's worth a demo, but budget for AI credits as well as seats, and check where call recordings are stored.

Sourcesx.com post ↗ · dreamteam.ai ↗

Matt Pocock's one-year history of AI coding: no-one writes code, but reading it still matters

@mattpocockuk posted a month-by-month history of AI coding since November 2025: Ralph loops in January, an explosion of shareable skills in March, 'loop engineering' hype in June, a 'software factory' in every conference talk by July, Grok Bot and long-running manager threads in September, and promised speed gains still missing in October. His summary: no-one writes code any more, but reading it still seems essential. It's one well-known developer's view rather than data. Disclosure: Grok Bot is the tool I use to put this digest together.

Why it mattersIf you pay for AI coding help or build with agents, keep budget for people who review and understand the code. Writing code got cheap this year; judgment didn't.

Sourcesx.com post ↗

Greg Isenberg: open-source your software as agent skills

@gregisenberg says open-sourcing is the new SaaS playbook: turn your best workflows into agent skills, publish them free on GitHub, make them work in Claude Code, Codex, Cursor and OpenClaw from day one, and ship an MCP server so agents can use your product. Then charge for what the free skills can't do: running every week with your data, hosting, support, team plans and done-for-you. In his words, free skills are the new free trial. It's his opinion, with no numbers attached.

Why it mattersIf you sell software or expertise, packaging your method as free skills puts your brand inside customers' AI tools every day. Keep your data, hosting and support as the part people pay for.

Sourcesx.com post ↗

How we knowCompany figures are labelled as their own claims; anything we couldn't verify is labelled as a report or a claim.