New AI models of November 2026: what changes with Claude Fable 5.2, Gemini 4 Carbon and the GPT-6 family
Three new AI models arrived in one week. We compare prices and test results for Claude Fable 5.2, Gemini 4 Carbon and GPT-6 using each company's official data.
· 4 min read
In just six days, three new AI models arrived from the industry's leading companies: OpenAI launched GPT-6 Cyber on November 5, Anthropic introduced Claude Fable 5.2 and Mythos 5.2 on November 9, and Google released Gemini 4 Carbon on November 10. Each targets something different: programming and agent work, business task automation or cybersecurity.
This comparison relies only on each company's official announcements and pricing pages. We have no commercial relationship with any of them.
3
Frontier models launched
between November 5 and 10
US$2
Lowest input price
per million tokens, Gemini 4 Carbon (introductory)
2
Models with restricted access
Mythos 5.2 and GPT-6 Cyber
Source: Anthropic · Google · OpenAI, API pricing
View the data as a table
| Frontier models launched | 3 | between November 5 and 10 |
| Lowest input price | US$2 | per million tokens, Gemini 4 Carbon (introductory) |
| Models with restricted access | 2 | Mythos 5.2 and GPT-6 Cyber |
What each of the new AI models brings
| Model | Company | Launch | Strengths | Who can use it |
|---|---|---|---|---|
| Claude Fable 5.2 | Anthropic | Nov 9 | Agentic coding, science and office work | Paid Claude plans and API, also on AWS, Google Cloud and Azure |
| Claude Mythos 5.2 | Anthropic | Nov 9 | Cybersecurity and life sciences | Vetted organizations only |
| Gemini 4 Carbon | Nov 10 | Coding, research and business agents | API for paying developers; Gemini app and AI Mode for AI Ultra subscribers | |
| GPT-6 Cyber | OpenAI | Nov 5 | Authorized vulnerability discovery and penetration testing | Vetted customers in the Daybreak Red program |
Claude Fable 5.2 is Anthropic's new flagship model. Mythos 5.2 is the same model without certain cybersecurity and biology safeguards, which is why it is only provided to vetted organizations, according to the company.
Gemini 4 Carbon is Google's frontier model for coding, research and agent tasks. Google highlights that it ranked first in external tests such as Zapier's AutomationBench, at more than two times lower cost per task than the next best model, according to its announcement.
GPT-6 Cyber is a cybersecurity-focused version of GPT-6 Astra, OpenAI's most capable model. OpenAI rated it "Critical" under its cybersecurity preparedness framework. On its ExploitGym test it scored 48.5%, compared with 42.4% for GPT-6 Astra and 32.1% for GPT-5.6 Cyber, according to eSecurity Planet, citing OpenAI's announcement.
How much the new AI models cost
Companies charge developers per million tokens, the units models split text into: one million tokens equals about 750,000 English words. Output tokens, meaning what the model writes, cost more than input tokens:
- Input
- Output
Source: Anthropic, Google and OpenAI, standard prices as of Nov 11, 2026
View the data as a table
| Input | Output | |
|---|---|---|
| Gemini 4 Carbon (through Nov 30) | 2 | 10 |
| Gemini 4 Carbon (from December) | 4 | 20 |
| Claude Fable 5.2 | 8 | 40 |
| GPT-6 Astra | 10 | 50 |
| GPT-6 Cyber | 20 | 100 |
- Gemini 4 Carbon has an introductory price of US$2 and US$10 that doubles after November 30, 2026.
- Claude Fable 5.2 and Mythos 5.2 cost the same; fast mode costs twice as much (US$16 and US$80).
- OpenAI offers cheaper models in the GPT-6 family, such as GPT-6.1 Sol (US$2 and US$10) and GPT-6 Luna (US$0.10 and US$0.50).
For people who use these models in a chat app, such as ChatGPT, Gemini or Claude, the cost depends on the subscription plan, not on tokens.
How they compare in tests
The only table comparing the three main models on the same tests is the one Anthropic published in its announcement. That is why it should be read with caution: it was produced by one of the competing companies, and each lab picks the tests where it performs best.
- Claude Fable 5.2
- GPT-6 Astra
- Gemini 4 Carbon
Source: Anthropic, Claude Fable 5.2 announcement (Nov 9, 2026)
View the data as a table
| Claude Fable 5.2 | GPT-6 Astra | Gemini 4 Carbon | |
|---|---|---|---|
| Agentic coding (FrontierSWE v2) | 69.9% | 65.5% | 61.2% |
| Business workflows (AutomationBench) | 51.4% | 41.4% | 53.9% |
| Computer use (OSWorld 2.1) | 89.3% | 79.5% | 81.5% |
| Chart reading (Chartography) | 73.4% | 71% | 72.5% |
The results show that no model wins at everything: Fable 5.2 leads in coding and computer use in that table, while Gemini 4 Carbon scores best in business workflow automation, which matches what Google reports. In chart reading, the three are within three points of each other.
What changes with the new AI models
Three trends repeat across the November launches:
- Agents over chat. The tests companies highlight no longer measure answers to questions but complete tasks: coding, using a computer or automating processes.
- Specialized models with controlled access. Mythos 5.2 and GPT-6 Cyber show that cybersecurity capabilities are being provided only to vetted users because of the risk of misuse.
- Price competition. Google launched Carbon at half its final price, and OpenAI offers lighter versions at a fraction of the cost of its main model.
To compare assistants from a user's point of view, read our comparison of AI virtual assistants. If you are interested in programming, see the AI tools for code generation and how much code AI already writes in our article on AI code generation.
New AI models change almost every month. We will update this comparison when independent tests are available or prices change.
What it means for you
- If you use AI at work: there is no single best model for everything. Try the same task in two or three assistants before choosing a paid plan.
- If your company builds with AI: compare cost per task, not just price per token, and keep in mind that the Gemini 4 Carbon promotion ends on November 30.
- If you work in IT security: specialized models require prior verification; if your organization needs them, start that process early.
Sources
- new AI models
- Claude Fable 5.2
- Gemini 4 Carbon
- GPT-6
- language models