General availability arrived November 10, 2026, with access spread across multiple Google platforms. Paying developers can reach Gemini 4 Carbon through the Gemini API in Google AI Studio. Enterprises have access through Gemini for Business and specialized agents including CodeMender and AlphaEvolve.
Consumer access is limited to Google AI Ultra subscribers via the Gemini app, AI Mode in Search, and Antigravity. The cited Google announcement does not specify whether Carbon will reach lower-priced tiers or free users.
The Introductory Price Makes a Real Case Against Claude Fable
Through November 30, Carbon’s token rates are roughly one-quarter of Claude Fable 5.2’s published standard price.
During the introductory period, Carbon costs $2 per million input tokens and $10 per million output tokens. Claude Fable 5.2 is listed at $8 per million input tokens and $40 per million output tokens, with an optional fast mode doubling those rates.
A Pricing Cliff Arrives December 1
Carbon’s rates double on December 1, landing it exactly where Claude Opus 5.5 already sits.
After November 30, Carbon’s standard rates rise to $4 per million input tokens and $20 per million output tokens. At that point, Carbon and Claude Opus 5.5 carry identical published prices, and the gap with Fable 5.2 narrows from fourfold to twofold.
Benchmark Data Shows a Mixed Picture
For latency-sensitive workloads, the gap is significant: Carbon averages 26.1 minutes per task while Fable 5.2 completes the same tasks in 16.9 minutes at high effort.
Artificial Analysis’s Intelligence Index scores Carbon at 60 and Fable 5.2 at 62, both at high effort. Carbon’s estimated average cost per Intelligence Index task is $7.76, against $9.08 for Fable at high effort, making Carbon the cheaper option at that configuration despite the modest performance gap.
Effort settings shift the comparison further. Artificial Analysis reports Fable at low effort costs $3.10 per task, against $6.59 for Carbon at medium effort, though those configurations are not directly comparable. Token rates alone do not resolve the question.
Task Type Determines Which Model Leads
Carbon dominates terminal benchmarks, while Fable holds the advantage on repository-level coding tasks.
On Artificial Analysis’s Terminal-Bench 4.0, Carbon scores 65.2% against Fable’s 59.6%. Fable reverses that result on FrontierCode-AA 2.0, scoring 48.8% to Carbon’s 40.6%. The better model depends on what a given workflow actually requires.
Total Workflow Cost Is the Right Metric
Thinking tokens, retries, and reviewer time all factor into what Carbon actually costs to run.
Google’s pricing documentation confirms that billed output includes thinking tokens and that Carbon is unavailable on the free tier. The cited Google announcement does not mention downloadable weights or a self-hosting license.
Teams evaluating Carbon need to account for token consumption across reasoning steps and task failure rates. Latency against their SLAs and the human time spent reviewing or correcting outputs add further cost.
Run the Pilot Before November 30
The introductory window is the right time to generate real cost data, not rely on marketing comparisons.
A controlled pilot using identical tasks, tools, permissions, and acceptance criteria will produce more useful numbers than published benchmarks alone. For terminal automation and business workflows that supercharge your productivity, Carbon’s benchmark results are competitive or stronger; for repository-level coding tasks, Fable’s FrontierCode-AA lead is worth weighing carefully.
Every estimate should model both the current $2/$10 rate and the December 1 rate of $4/$20 per million input and output tokens. A workflow that looks cheap today may price differently in five weeks.
