By ZOIPI Editors 6 min read

AI News for Builders: The Big Model Launches of Mid-November

Five launches from the past ten days that matter if you build things: Claude Fable 5.2, Gemini 4 Hydrogen, ElevenLabs Scribe v3, GPT-6.2 Astra and Sol, and Microsoft's MAI Code 2. What each one is for, in plain words.

If you blinked during the second week of November, you missed about five model launches. The short version: Anthropic, Google, OpenAI and Microsoft all shipped new models between 9 and 17 November, and ElevenLabs put out a new speech-to-text model in the middle of it. Most of them are about coding and agents. One is about turning audio into text.

Here's what each one is, who it's for, and whether you should care this week. Every item links to the company's own announcement. We haven't run these models ourselves, so wherever there are claims about performance, we say whose claim it is.

Claude Fable 5.2 (Anthropic, 9 November)

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, and the funny part is that, by Anthropic's own account, they're the same model. The difference is who gets in. Fable 5.2 is open to everyone through the API and on AWS, Google Cloud and Microsoft Azure. Mythos 5.2 is only for verified cybersecurity and life-sciences organizations, plus Claude Enterprise customers through Claude Security.

What's it for? Anthropic pitches Fable at hard, open-ended technical work. It also says, refreshingly, that most everyday coding and knowledge work should start with Opus 5.5, and that you should move up only when a task needs it. Fable 5.2 can also hand simpler jobs to smaller models. The announcement lists the price at $8 per million input tokens and $40 per million output, which is cheaper than Fable 5.1.

If you use Claude or Claude Code day to day, the practical takeaway is simple. Keep your default where it is, and reach for Fable when something keeps falling over.

ElevenLabs Scribe v3 (10 November)

This one's for anyone who turns calls, podcasts or voice notes into text. ElevenLabs released Scribe v3 and Scribe v3 Live. The first handles recorded audio and the second handles live audio.

According to ElevenLabs, the main upgrades are fewer word errors than Scribe v2, better speaker labeling on recordings with three or more people, and support for up to 1,000 "keyterms" (your product names, people's names, jargon). The live model gets speaker labeling for the first time. It also uses the conversation itself, not just silence, to tell when someone has finished talking, which matters if you're building a voice agent that keeps interrupting people.

Both are in the ElevenLabs API now. If you already use ElevenLabs for voices, this keeps the whole audio loop in one place. And if you're shopping for transcription more broadly, our roundup of AI transcription tools covers the ready-made apps.

Gemini 4 Hydrogen (Google, 12 November)

Google's month went in two steps. First came Gemini 4 Carbon, the big reasoning model for hard multi-step problems. Then on 12 November it released Gemini 4 Hydrogen, the fast, efficient one, for everyday coding and agent workflows: working in a repo, debugging, web dev and multimodal tasks.

The interesting bit for builders is that Hydrogen is meant to work alongside Carbon and Argon rather than replace them. Google describes it running on its own at four thinking levels or as part of a team with the bigger models. That's a pattern you're now seeing everywhere: a cheap, quick model does most of the work and passes the hard parts up.

You can build with it through the Gemini API in Google AI Studio and Android Studio, in Gemini for business, and in Google Antigravity for paid AI Pro or Ultra subscribers. Google's post doesn't list a price for Hydrogen, so we won't guess. If you mostly use the Gemini app, you won't need to do anything.

GPT-6.2 Astra and Sol (OpenAI, 17 November)

OpenAI shipped two models at once. OpenAI's announcement calls Astra its frontier model for your hardest projects, aimed at professional work and research. Sol is the efficient, versatile one for everyday work and multi-step agent workflows.

There's one catch worth knowing before you go looking. According to OpenAI's announcement thread, both are rolling out in ChatGPT Work and Codex on eligible plans (Plus, Pro, Enterprise and Edu, depending on the plan) and in the API, but not in regular ChatGPT chat. So if you open ChatGPT and don't see them, nothing is broken.

Pricing per million tokens is listed at $8 in and $40 out for Astra, and $2 in and $10 out for Sol. Yes, Astra and Fable 5.2 have the same list price. Make of that what you will.

MAI Code 2 and six friends (Microsoft, 17 November)

At Ignite, Microsoft announced seven new models in one go. For builders, the headline is the coding pair. MAI Code 2 is the flagship coding model and MAI Code 2 Flash is the cheap, fast one for routine fixes, small edits and quick explanations. Both are live in GitHub Copilot now.

The rest of the list covers security (MAI Cyber 2 Flash), images (MAI Image 2.7), live speech-to-speech (MAI Realtime 1), medical transcription in Dragon Copilot, and Decision 1.1, a model whose job is to pick which model should answer. Microsoft makes plenty of price-performance claims against competitors. Those are Microsoft's numbers, and we'd wait for people to try them before repeating them.

Microsoft also introduced "MAI Licensed Weights", which let a company customize one of these models with its own data and run it inside its own Azure setup. That's mostly for big companies today, but it's a sign of where things are going.

What this means if you're building something

A few honest patterns from all of this:

  • Two sizes is the new normal. Almost every launch came as a pair: a big, careful model and a fast, cheap one. Default to the cheap one and move up only when you have to. Anthropic and Google both said this out loud.
  • Coding agents are the main event. Four of the five launches lead with coding or agent work. If you build with AI, your editor and agent tools will keep changing under you, so don't get too attached to one setup.
  • Where a model lives matters as much as how good it is. Astra isn't in regular ChatGPT, Mythos needs verification and MAI Code 2 is in Copilot first. Check access before you plan around a model.
  • Benchmarks are marketing until proven otherwise. Every announcement had a chart where the author's model wins. Run your own task.

The easiest way to do that last one is with a tool that sends the same prompt to several models. Our Multi-LLM Aggregators category has a bunch of them, and the OpenAI listing collects the official ways in to OpenAI's models.

Bottom line

It was a busy ten days, but you don't need to switch anything today. If you code with AI, try the cheaper new model inside the tool you already use: Sol in Codex, MAI Code 2 in Copilot or Hydrogen in AI Studio. Keep the expensive ones for the problems that beat it. If you work with audio, Scribe v3 is the one to look at. Everything else can wait until real users have had a week with it.