TL;DR
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
Google made Gemini 4 Carbon available to paying developers, businesses and Google AI Ultra subscribers on November 10, alongside a model card. On the new MKA benchmark, Carbon scored 54.3%, behind Claude Fable 5.2 at 56.6%; Google DeepMind reported a lower cost per task for Carbon, though the benchmark was co-developed by Google and its tasks are private.
Google opened Gemini 4 Carbon to paying developers, businesses and Google AI Ultra subscribers on November 10, the same day Google DeepMind and Artificial Analysis released results from their new MKA benchmark for AI agents. Carbon scored 54.3%, placing second behind Anthropic’s Claude Fable 5.2 at 56.6%; Google DeepMind reported that Carbon cost $5.63 per MKA task, compared with $35.28 for Fable 5.2.
Carbon is available through the Gemini API and Google AI Studio for developers, and through Gemini for business. In the consumer Gemini app and AI Mode in Search, access is limited to Google AI Ultra subscribers. Google’s model card identifies Carbon as a top-tier model built on Gemini 4 Argon and aimed at demanding work, including complex coding and open-ended scientific research.
Google describes Carbon as a model that spends more tokens on reasoning than Argon, including extended verification. That approach can use more tokens and may be slower. Google lists hallucinations, occasional slowness or timeouts, and higher token use at elevated thinking levels among its known limitations. The company says Argon remains the better fit for most software engineering and knowledge work.
Carbon’s introductory API price is $2 per million input tokens and $10 per million output tokens through November 30. Google says those rates will rise to $4 and $20, respectively, on December 1. The MKA cost comparison is a separate measure: Google DeepMind reported a per-task cost, rather than a price per million tokens.
Cost and Performance in Agent Tasks
The MKA result presents a practical trade-off for organizations evaluating models that must interpret information in formats such as blueprints, scans and charts. Carbon finished close to the top score while its reported cost per task was substantially lower. That could matter for teams running many specialized tasks, although the published comparison alone does not establish what any particular customer will pay or how the models will perform on its own workload.
The ranking also shows why one benchmark cannot settle which model is best overall. MKA measures a particular kind of agent work, while other evaluations test coding, computer operation or broad reasoning. Carbon’s second-place finish on MKA and its second-place standing on the current Artificial Analysis Intelligence Index are useful reference points, but buyers still need to weigh task fit, reliability, latency and pricing.
For everyday users, availability depends on the product and subscription. Developers and businesses can access Carbon through paid services, while use in Google’s consumer app and Search AI Mode is restricted to Ultra subscribers. Google has not presented broad free access as part of this release.
Products Worth Considering
DJI Mini 4K Drone with 4K UHD Camera
Logitech MX Mechanical Wireless Illuminated Keyboard, Tactile Quiet, Graphite
Logitech MX Ergo S Advanced Wireless Trackball Mouse USB-C – Graphite
As an affiliate, we earn on qualifying purchases.
Carbon’s Scores Across Benchmarks
Google’s model card reports Carbon leading on 15 of 17 benchmarks selected for its comparison, with results dated as of October 2026. The comparison includes Gemini 4 Argon, OpenAI’s GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1. It does not include Claude Fable 5.2, which the source material says was released on November 9.
The Google-selected results include narrow leads: Carbon scored 9.6% on SWE-sweep against GPT-6 Astra’s 9.5%, and 70.4% on the Vals Index against Claude Opus 5.5’s 70.2%. On other tests, the gaps were wider. Carbon scored 96.2% on GraphWalks, which tests reasoning across very long inputs, while the best rival in that comparison scored 71.8%. On Zapier’s AutomationBench, Carbon scored 53.9% against 42.5%.
Carbon did not lead every test. GPT-6 Astra scored 65.5% on FrontierSWE 2, compared with Carbon’s 61.2%; Claude Opus 5.5 scored 87.2% on OSWorld-2.1, where Carbon scored 81.5%. Artificial Analysis’ current Intelligence Index version, v5.0.1, gives Carbon at its high setting a score of 60. Fable 5.2 scores 65 at its maximum setting and 62 at high; Claude Opus 5.5 scores 59 and GPT-6 Astra 57.
Carbon had earlier been made available to a small group of cyber defenders through Google’s Fairwind Program, with Artificial Analysis listing October 28 as its release date. Google says enhanced security features remain limited to verified defenders.
““tasks that test the limits of a model’s intelligence””
Limits of the MKA Comparison
MKA covers 200 expert-built tasks across 10 sectors, but Google DeepMind and Artificial Analysis have not made the tasks public, according to the source material. That prevents outside researchers from independently checking the task set or reproducing the published comparison. Google DeepMind co-developed the benchmark, and Carbon performs well on it, so the results warrant attention to the benchmark’s design and independent replication.
The source says the top three MKA scores are within the margin of error, but does not provide the underlying uncertainty range here. It is consequently unclear how confidently the 2.3 percentage-point gap between Fable 5.2 and Carbon separates their performance. The reported task costs also come from Google DeepMind’s report; the supplied material does not specify all assumptions behind those estimates.
Google’s model card and Artificial Analysis index use different sets of tests and model settings. Their scores should not be treated as directly interchangeable. Carbon’s actual cost and performance for a customer will depend on usage, selected thinking level and the task being run.
Pricing Changes in December
Google’s announced API pricing is scheduled to change on December 1, 2026, rising from $2 to $4 per million input tokens and from $10 to $20 per million output tokens. Developers comparing Carbon with other models will need to account for that date, as well as their expected input and output volumes.
Further clarity would come from independent evaluations using public or reproducible tasks, fuller reporting of the MKA cost assumptions, and results on real customer workloads. For now, the published evidence establishes Carbon’s availability, its reported benchmark scores and Google DeepMind’s cost estimate; it does not establish a universal performance or savings advantage.
Key Questions
Who can use Gemini 4 Carbon?
Paying developers and businesses can access it through the Gemini API, Google AI Studio and Gemini for business. In the Gemini app and AI Mode in Search, access is for Google AI Ultra subscribers.
How did Carbon score on MKA?
Carbon scored 54.3%, behind Claude Fable 5.2 at 56.6% and ahead of Claude Opus 5.5 at 53.0%. The source says the top three results are within the margin of error.
How much does Carbon cost?
Through November 30, Google’s API price is $2 per million input tokens and $10 per million output tokens. From December 1, the listed rates rise to $4 and $20.
Is Carbon the top-ranked model?
Not on the reported MKA results: Claude Fable 5.2 scored higher. On Artificial Analysis’ Intelligence Index v5.0.1, Carbon at high scores 60, below Fable 5.2 at high with 62.
Can outside researchers verify the MKA results?
The tasks are described as private, so outsiders cannot directly inspect or reproduce the full benchmark from the information provided. Google co-developed MKA; independent replication would help clarify how well its scores generalize.
Source: Thorsten Meyer AI
