November 11, 2026

Google said on 10 November 2026 that Gemini 4 Carbon is available through the API in Google AI Studio and in Gemini for business, for all paying developers and enterprises. The price is $2 per million input tokens and $10 per million output tokens. A footnote says $4 and $20 apply after the introductory period expires on 30 November 2026.

The catch: The price per token is not the bill. Google says Carbon uses more tokens than its Argon model, and that Argon "remains the ideal choice for the majority of software engineering and knowledge work". The post gives no token count per task. In the Gemini app, Google says Carbon is rolling out to Google AI Ultra subscribers.

What to do: Budget any Carbon pilot at $4 and $20. Run the same 50 tasks on Argon and Carbon and record tokens and dollars per completed task. Move a workload only where Carbon finishes tasks Argon fails.

Microsoft Research published a method on 10 November 2026, called Lazarus, that plants a backdoor in an open-weight model by rewriting fewer than 10 rarely used neurons, with no training. The paper reports that seven edited neurons in NVIDIA's Nemotron-Terminal-8B produced the attacker's command on at least 98.8% of triggered test trajectories. Microsoft says the model passes standard evaluations until a trigger appears, such as a date.

Why it matters: The paper notes that the weights people run are often a quantised, fine-tuned or merged copy from a public hub, not the original release. Microsoft says an agent runs the command with whatever access you gave it.

The catch: The paper's detector flagged 91.4% of edited test checkpoints and 9.7% of clean ones. Microsoft says no scan can prove a model safe, and that reliable detection needs a comparison with the original weights.

What to do: Record the source and file hash of every open-weight checkpoint you run, and compare each with the publisher's original. Keep credentials out of the agent's reach, and require approval before a generated command runs.