- What did Google release on November 10?
- What does Gemini 4 Carbon cost in Canada?
- Where does Carbon lead on Google’s own benchmarks?
- What is the catch for a Canadian buyer?
- What test should you run before switching?
Google made Gemini 4 Carbon generally available on November 10, 2026, and called it the strongest model in its Gemini 4 family. Carbon targets long-horizon agent work, complex coding and scientific reasoning. The price is easy to read. The cost per finished task decides whether your firm switches.
The direct answer: Gemini 4 Carbon costs $2.84 CAD per million input tokens and $14.18 CAD per million output tokens at the introductory API price, converted at the Bank of Canada rate for November 10, 2026. Google ranks Carbon 1st on 3 workflow benchmarks and states it uses more tokens than Argon to work through hard problems. Buy it for the hardest tasks in your queue, price it by the finished task, and keep Argon beside it for ordinary work.
What did Google release on November 10?
Carbon spent 2 weeks with cybersecurity defenders in Google’s Fairwind Program before this wider release. Google built it for tasks that test the limits of a model, such as multi-step coding and debugging and open-ended research, and it uses more tokens than Gemini 4 Argon to explore, reason and run verification loops. A preview feature in Antigravity, called /swarm, turns 1 prompt into a team of agents a developer can track. Google states plainly that Argon remains the ideal choice for the majority of software engineering and knowledge work.
What does Gemini 4 Carbon cost in Canada?
The introductory API price for Gemini 4 Carbon, per million tokens
| Measure | US price | Canadian price |
|---|---|---|
| Input tokens | $2 | $2.84 CAD |
| Output tokens | $10 | $14.18 CAD |
The Canadian figures convert Google’s published US prices at US$1 = CAD 1.4181, the Bank of Canada daily rate for November 10, 2026. The token is the unit on the invoice, and the unit misleads in the way this magazine flagged in Stop counting tokens. Even the man who sells them calls the metric fake. A model that spends more tokens per task can cost more per finished task at a lower token price. Routing work between a pricey reader and a cheap one is the other half of this math, priced in Spotify’s plugin cut Claude Code token use by 90% in its own test. Your token bill is a settings problem.
Where does Carbon lead on Google’s own benchmarks?
Google ranks Carbon 1st on CursorBench 4.0, on Perplexity’s WANDR and on Zapier’s AutomationBench, at more than 2x lower cost per task than the next best external model on those tests. These are Google’s published results, read as vendor claims. The independent picture on this site is colder. In The best AI in the world scored 12% on real professional work, Carbon scored 10.9% on real professional tasks in the GDPx test, at a cost per finished task this magazine calculated near $211 CAD. Both readings can stand at once. Carbon leads its maker’s workflow tests, and it still fails most real professional work.
What is the catch for a Canadian buyer?
Access splits by surface. Paying developers and enterprises reach Carbon through the API in AI Studio and through Gemini for business. Everyday access in the Gemini app and AI Mode in Search is rolling out to Google AI Ultra subscribers. The cybersecurity build runs through the Fairwind Program, by application. Google reports a 0% attack success rate on 1 prompt injection benchmark, with its own caveat that prompt injection remains an area for improvement. Read that result as a vendor’s test log. No hands-on test ran for this post.
What test should you run before switching?
Take 5 of your hardest recurring tasks. Run each task on Carbon and on the model your team uses today. Record the finished tasks, the failed tasks and the full token spend. Divide spend by finished tasks. The model with the lower cost per finished task earns the work, whatever its token price says.
Carbon is a specialist at a generalist’s shelf price. Specialists earn their fee on the hard cases, and the hard cases are the only fair test.
