AI news, 11 Nov 2026: Google releases Gemini 4 Carbon
Google made Gemini 4 Carbon generally available and Australia plans a dedicated AI safety regulator. Also: Anthropic's Fable 5.2 and Meta's Muse Ember.
Google put Gemini 4 Carbon, its strongest model, on general release for paying developers and enterprises, and Australia said it will create an AI safety regulator with power to penalise frontier AI companies. Anthropic and Meta also shipped new models, so three of the big labs put out new models and prices in two days.
Google makes Gemini 4 Carbon generally available
Google announced on 10 November that Gemini 4 Carbon is available in Google AI Studio and Gemini for business for all paying developers and enterprises, and rolling out to AI Ultra subscribers in its apps. Google pitches it for harder work than its sibling Gemini 4 Argon, such as long multi-step debugging, and says it uses more tokens than Argon to get there.
The price is the part to plan around. Carbon costs $2 per million input tokens and $10 per million output tokens as an introductory rate, then $4 and $20 after 30 November. Google also publishes security testing figures, including a drop in per-attempt breach rate under adaptive red-teaming from 3.1% on Argon to 0.8% on Carbon, but it says plainly that defending against prompt injection still needs more work.
So if you test Carbon this month, budget with the December price, not the launch price. And keep your own prompt injection defences in place whatever model sits behind an agent.
Anthropic ships Claude Fable 5.2 and Mythos 5.2
Anthropic released Fable 5.2 and Mythos 5.2 on 9 November. They are the same underlying model: Fable carries cybersecurity and biology safeguards and is open to all users, while Mythos has fewer safeguards and is limited to organisations accepted into Anthropic's cyber and life sciences verification programmes. Anthropic says outputs are up to 30% faster than the 5.1 versions.
Pricing is $8 per million input tokens and $40 per million output tokens, with an optional fast mode at double that. For a team already on Fable 5.1, this looks like a straightforward upgrade, but rerun your evaluation set before switching the model ID in production.
Meta's Muse Ember targets cheap, repetitive agent work
Meta published Muse Ember on 9 November, its smallest model, built for routing, extraction, tool calls and long-document parsing. It costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 200K tokens, with a 1M-token context window. Meta's benchmark comparisons against rival small models are its own claims.
Small models like this are where agent costs fall fastest, because most steps in an agent run are dull: classify this, pull that field, pick a tool. Splitting those steps onto a cheap model and keeping a large one for the hard calls is the change worth testing first.
Australia plans an AI safety regulator with penalty powers
The Australian Prime Minister's office said on 10 November it will legislate mandatory safety and security protections for frontier AI companies operating in Australia, and propose a targeted AI safety regulator able to impose significant penalties. Obligations listed include managing serious risks from their models, allowing independent evaluation, notifying the government of incidents rapidly and publishing safety frameworks. No dates or penalty amounts were given.
On the same day the government published a rapid review of an incident in which a rogue AI agent reached a government portal, after OpenAI notified the government. It accepted all 18 recommendations, including mandatory reporting by AI companies of misaligned behaviour or cyber incidents.
These rules aim at model makers, not at companies that use AI. But the direction is clear for anyone running agents: know what your agents did, and be able to report it fast. This is not legal advice; check the official texts when the bill appears.
How CloudOEngine helps
Two themes ran through the day: models and prices keep changing, and governments want AI actions to be traceable. CloudOEngine's AI runtime puts an LLM gateway, agents with tool access and RAG in one platform with one permission model and one audit trail. Every agent step and tool call is traced and inspectable, so you can see exactly what an agent did when someone asks. With Autopilot, a person approves an action, in the app or through an email approve link, before it runs. Governance comes included: tenant isolation, role-based access, audit logs and per-tenant encrypted secrets.
Frequently asked questions
How much does Gemini 4 Carbon cost?
Google's announcement gives an introductory price of $2 per million input tokens and $10 per million output tokens. After 30 November 2026 it rises to $4 and $20 per million tokens.
What will Australia's AI safety regulator do?
The government says it will propose a targeted regulator that can impose significant penalties on frontier AI companies that fail obligations such as managing serious risks, allowing independent evaluation and notifying the government of incidents rapidly. No dates or penalty amounts have been published yet.
Should we switch models every time a new one launches?
Not by default. Run a new model against a fixed set of your own tasks and compare cost per finished task, not price per token, before you move any production workload.