Da Nang, VN,--:--:--
All articles
  • #ai
  • #llm

Nine days, three frontier launches: the November 2026 AI model wave

Claude Fable 5.2, Gemini 4 Carbon and Hydrogen, GPT-6.2 Astra and Sol — plus Anthropic's IPO. What actually shipped, what it costs, and what it changes for developers.

NDNgoc Duc4 min read

Between November 9 and November 17, 2026, Anthropic, Google and OpenAI each shipped a new frontier model — and Anthropic went public the same week. If you build with LLM APIs, here's the condensed version.

The launches

Claude Fable 5.2 and Mythos 5.2 — Anthropic, Nov 9

Anthropic describes Fable 5.2 and Mythos 5.2 as the same underlying model. Fable ships to everyone with safeguards that block high-risk cybersecurity and biology requests; Mythos has no such limits and is only available to vetted organisations through verification programs.

  • API ID: claude-fable-5-2
  • Price: $8 input / $40 output per million tokens, with an optional fast mode (up to 2.5× faster) at $16 / $80
  • Where: Claude apps, Claude Code, the Claude Platform, plus AWS, Google Cloud and Microsoft Azure
  • Headline claims: up to 30% faster output than 5.1, about 55% fewer safeguard blocks, and 49.1% on the FrontierCode 2.0 agentic coding benchmark (up from 32.4% for Fable 5.1)

Interestingly, Anthropic itself still recommends Opus 5.5 as the everyday default and Fable for the hardest work.

Gemini 4 Carbon and Hydrogen — Google, Nov 10 and Nov 12

Google is rolling Gemini 4 out as a family:

  • Carbon (Nov 10) is the strongest model, aimed at cybersecurity and long-horizon agentic work. It's on the paid Gemini API tier in AI Studio and Android Studio, and in Google's agent-first IDE, Antigravity, for AI Ultra subscribers. The launch post doesn't list a price.
  • Hydrogen (Nov 12) is the cheap, fast "daily driver": gemini-4-hydrogen, $1.50 / $7.50 per million tokens, priced flat across a 1M-token context, with low/medium/high/xhigh thinking levels. Google reports 57% on Terminal-Bench 4.1, against 18.2% for Gemini 3.8 Flash.

Google also pitches pairing them: Carbon plans and coordinates while Hydrogen runs worker and verifier agents.

GPT-6.2 Astra and Sol — OpenAI, Nov 17

  • Astra — gpt-6.2-astra, $8 / $40 per million tokens, 20% cheaper than GPT-6 Astra. Pitched at the hardest professional and scientific work.
  • Sol — gpt-6.2-sol, $2 / $10, same price as GPT-6.1 Sol. OpenAI says it nearly matches Astra on merge-ready code changes at about a fifth of the cost.

Both are in the API and in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu — but not in regular ChatGPT chat. OpenAI says Astra is the first model it has deployed externally that crosses the "High" threshold for AI self-improvement in its Preparedness Framework, and both launch with tighter cybersecurity safeguards. Early reviews were mixed: some testers reported loops and tasks declared "done" too early, and OpenAI said significant changes landed in the final days before launch.

Side by side

ModelAPI IDInput / Output per 1M tokens
Claude Fable 5.2claude-fable-5-2$8 / $40
GPT-6.2 Astragpt-6.2-astra$8 / $40
GPT-6.2 Solgpt-6.2-sol$2 / $10
Gemini 4 Hydrogengemini-4-hydrogen$1.50 / $7.50
Gemini 4 Carbonnot stated in launch postnot stated

And then Anthropic went public

On November 18, Anthropic priced its IPO at $83 per share. It started trading on Nasdaq under ANT the next day and, after a volatile session, closed at $97.17 — about 17% above the offer price — while the Nasdaq itself fell. Coverage of how much was raised varies by outlet (figures range from about $81 billion to more than $100 billion), so I'll wait for the final filings before quoting one.

What this means if you build with LLMs

  1. Two tiers is the new normal. Every lab now has a premium "think hard" model and a cheap workhorse. Default to the workhorse and escalate only when a task fails.
  2. Price per token isn't price per task. All three vendors argue their new models finish work with fewer tokens. Measure cost per completed task, not per million tokens.
  3. Orchestration is becoming a product feature. Anthropic (Fable advising or delegating to smaller Claude models) and Google (Carbon + Hydrogen in Antigravity) both ship "big model plans, small model executes" patterns. It's worth designing your agents this way too.
  4. Safety gating affects real workflows. If you do security work, expect refusals or routing on the public models, and check whether a verification program (Mythos, OpenAI's defender access) fits you.
  5. Keep the provider swappable. With three launches in nine days, a thin abstraction over model IDs and a small eval suite will pay for themselves quickly.

Sources

Got a project in mind?

Let's talk