🔍 Read the full analysis: Gemini 4 Carbon Takes Second On MKA For One-Sixth Of The Leader’s Price on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Google opened Gemini 4 Carbon to paying customers on November 10, the same day it and Artificial Analysis introduced the MKA benchmark for AI agents handling visual and technical documents. Carbon scored 54.3%, second to Claude Fable 5.2 at 56.6%, while Google DeepMind reported a lower MKA test cost: $5.63 per task versus $35.28. The benchmark was co-developed by Google, and its task set is private, limiting independent scrutiny.
Google opened its Gemini 4 Carbon model to paying developers, businesses and Google AI Ultra subscribers on November 10, the same day Google DeepMind and Artificial Analysis released a new benchmark for AI agents working with visual and technical documents. Carbon scored 54.3% on MKA, placing second behind Anthropic’s Claude Fable 5.2 at 56.6%; Google DeepMind reported Carbon cost $5.63 per task, compared with $35.28 for the leading model.
MKA evaluates agents on 200 expert-built tasks across 10 sectors, including work involving blueprints, scans and charts rather than only plain text. The published results put Claude Fable 5.2 first at 56.6%, Gemini 4 Carbon second at 54.3%, and Claude Opus 5.5 third at 53.0%. The benchmark report says the top three results fall within the margin of error, so the small score gap should not be treated as a definitive separation in capability.
The price comparison comes from Google DeepMind’s report, not an independently audited billing analysis. It lists Carbon at $5.63 per MKA task and Fable 5.2 at $35.28. Those figures describe costs on this benchmark and do not establish what every customer will pay for a real-world task, which can vary with input, output and model settings.
Carbon is available through the paid Gemini API, Google AI Studio and Gemini for business. In the Gemini app and AI Mode in Search, access is limited to Google AI Ultra subscribers. Google lists introductory API rates of $2 per million input tokens and $10 per million output tokens through November 30. From December 1, the listed rates rise to $4 and $20, respectively.
Gemini 4 Carbon takes second at a much lower reported task cost
Google’s model scored close to the MKA leader, while Google DeepMind reported a much lower cost per benchmark task. The test’s private tasks limit independent scrutiny.
Close results, different reported costs
The benchmark report places the top three within the margin of error.
Carbon is close to the leader
Its score trails Fable 5.2 by 2.3 percentage points, a gap the report says is within the margin of error.
Opus 5.5 scored 53.0%
The top three results are tightly grouped on this test.
What MKA tests
The benchmark evaluates AI agents on visual and technical documents.
MKA includes 200 tasks across 10 sectors.
Tasks involve materials such as blueprints, scans and charts.
Outside researchers cannot inspect or directly reproduce the prompts.
Access depends on the service
The November 10 release expanded paid access, with different terms across Google products.
- Developers: Gemini API and Google AI Studio
- Businesses: Gemini for business
- Gemini app and Search AI Mode require Google AI Ultra
- Some security features remain limited to verified defenders
API rates are scheduled to rise
Developers evaluating Carbon should account for the announced price change.
Lower Benchmark Cost, Close Scores
The results give companies evaluating AI agents a comparison that combines task performance and reported cost. Carbon’s second-place score is close to the leader’s, while its reported MKA task cost is substantially lower. For organizations testing agents on documents and visual information, that price difference may merit evaluation alongside accuracy, speed, reliability and the cost of human review.
It is not a general finding that Carbon is always cheaper or better value. Artificial Analysis’ separate Intelligence Index v5.0.1 gives Carbon a score of 60, compared with Fable 5.2 at 62 on its high setting and 65 at maximum. Artificial Analysis also reports that Carbon took 25.6 minutes per index task, against a 12.6-minute median for comparable models. That is a different test and a reminder that cost per task alone does not capture how quickly a system completes work.
For users, availability is also tiered. Developers and businesses can access Carbon through paid products, but use in Google’s consumer-facing app and Search AI Mode requires Google AI Ultra. That distinction matters when interpreting claims that the model is now broadly available: paid access has expanded, but access is not the same across Google’s services.
Popular right now
Ninja CREAMi Scoop & Swirl 13-in-1 Ice Cream and Soft Serve Maker, NC701
Shark Rotator Lift-Away Upright Vacuum with DuoClean PowerFins, Wine Purple, LA702
Shark NV752 Rotator Powered Lift-Away TruePet Upright Vacuum, Bordeaux
As an affiliate, we earn on qualifying purchases.
Carbon’s Release and Test Record
Carbon is Google’s top-tier model built on Gemini 4 Argon. Google describes it as intended for demanding work such as multi-step coding and open-ended scientific research. Its model card says Carbon uses more thinking tokens than Argon to examine hypotheses and run extended verification. Google also says Argon remains the better choice for most software engineering and knowledge-work tasks.
Google’s model card reports Carbon leading on 15 of 17 benchmarks selected by the company. That comparison reflects Google’s chosen tests and competitors, and its results are dated October 2026. The table does not include Claude Fable 5.2, released November 9. Carbon also does not lead every listed test: GPT-6 Astra scores higher on FrontierSWE 2, while Claude Opus 5.5 leads on OSWorld-2.1.
Before the wider paid release, Carbon was made available to a small group of cyber defenders through Google’s Fairwind Program. The source material says its stronger security features remain restricted to verified defenders. The November 10 release therefore marks a move into paid developer and business products, not unrestricted access to every capability.
““tasks that test the limits of a model’s intelligence””
— Google DeepMind, Gemini 4 Carbon model card
Limits of the MKA Comparison
MKA was co-developed by Google DeepMind, and the task set is private. That means outside researchers cannot inspect the questions or independently reproduce the reported scores from the published task materials. The benchmark offers a comparison, but those design and access limits should be weighed when interpreting Carbon’s result.
The score difference between Carbon and Fable 5.2 is 2.3 percentage points, and the report places the top three within the margin of error. The available material does not provide enough detail here to determine how results vary by sector, task type or repeated runs. It also does not establish whether the reported task costs include all relevant expenses for a customer deployment.
Google identifies hallucinations, occasional slowness or timeouts, and increased token use as Carbon limitations. The precise frequency of these issues in ordinary use is not stated. Nor do the benchmark results show how Carbon compares with Fable 5.2 across every workload or service setting.
Pricing Changes and Further Testing
Developers can test Carbon through Google’s paid API and AI Studio, while business customers can access it through Gemini for business. Google’s published API pricing is scheduled to move from $2 per million input tokens and $10 per million output tokens to $4 and $20 on December 1, 2026. Customers considering the model will need to account for that change and test performance on their own tasks.
Further evidence will depend on independent evaluations, additional benchmark disclosures and real-world reports on speed, reliability and cost. Until MKA’s private tasks or further validation are available, its rankings are best read as one comparison under a specific test design, rather than a complete measure of which model is best for every organization.
Key Questions
What was Gemini 4 Carbon’s MKA score?
Carbon scored 54.3% on MKA, placing second behind Claude Fable 5.2 at 56.6%. The report says the top three models were within the margin of error.
How much did Carbon cost per MKA task?
Google DeepMind reported a cost of $5.63 per task for Carbon, compared with $35.28 for Fable 5.2. These are reported benchmark costs, not a guarantee of customer costs on other workloads.
Who can use Gemini 4 Carbon?
It is available to paying developers through the Gemini API and Google AI Studio, and to businesses through Gemini for business. Access in the Gemini app and AI Mode in Search is limited to Google AI Ultra subscribers.
Is the MKA ranking independently verifiable?
Not fully from the information currently available. Google DeepMind co-developed MKA with Artificial Analysis, and the 200 task prompts are private, preventing outside researchers from directly checking or reproducing the test.
When does Carbon’s API price change?
Google lists introductory rates through November 30, 2026: $2 per million input tokens and $10 per million output tokens. The listed rates rise to $4 per million input tokens and $20 per million output tokens on December 1.
Source: ThorstenMeyerAI.com
