🔍 Read the full analysis: Why Gemini 4 Carbon’s Second-Place MKA Ranking Is Drawing Attention 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 Google DeepMind and Artificial Analysis introduced the MKA benchmark. Carbon scored 54.3%, placing second to Claude Fable 5.2 at 56.6%; the benchmark report says Carbon cost substantially less per task, though the top three scores are within the margin of error and Google co-developed the test.
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 announced a new benchmark for AI agents. Carbon scored 54.3% on MKA, placing second behind Anthropic’s Claude Fable 5.2 at 56.6%, while the benchmark report put Carbon’s cost at $5.63 per task versus $35.28 for Fable 5.2.
MKA is designed to test agents on tasks involving blueprints, scans and charts, rather than only plain-text prompts. Its 200 expert-built tasks span 10 sectors. Claude Fable 5.2 scored 56.6%, Gemini 4 Carbon 54.3%, and Claude Opus 5.5 53.0%. The source report says the scores of the top three fall within the benchmark’s margin of error, so the ordering should not be read as a definitive measure of a meaningful performance gap.
The cost figures come from Google DeepMind’s MKA report: $5.63 per task for Carbon and $35.28 for Fable 5.2. They are benchmark-specific figures, not a general price for every use of either model. Separately, Google’s model card reports Carbon leading on 15 of 17 selected benchmarks, but those results are based on tests and competing models chosen by Google and are dated October 2026. On two listed tests, GPT-6 Astra and Claude Opus 5.5 score higher.
Access depends on the product. Carbon is available through the paid Gemini API, Google AI Studio and Gemini for business; within the Gemini app and AI Mode in Search, it 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, rising to $4 and $20 from December 1, 2026.
Gemini 4 Carbon’s MKA result close score, lower reported cost
Carbon ranked second on a benchmark for agents working with visual documents and data. The close scores and benchmark-specific cost figures need careful interpretation.
What MKA evaluates
The benchmark tests agents on visual and structured materials, not only plain-text prompts.
Real-world document types
Tasks involve blueprints, scans and charts.
Broad task set
MKA includes 200 expert-built tasks across 10 sectors.
The top scores are close
The report says the top three results fall within the benchmark’s margin of error.
How to read the cost comparison
The reported amounts describe this benchmark evaluation, not every customer workload.
- Carbon paired a near-leading score with lower reported cost per task.
- Businesses evaluating document-focused agents may want to test it.
- The top-three ordering is not a definitive performance gap.
- Customer costs can vary with prompts, output length, retries and workload.
Access and pricing
Availability depends on the Google product.
Carbon is available through the paid Gemini API, Google AI Studio and Gemini for business.
Gemini app and AI Mode in Search access is limited to Google AI Ultra subscribers.
Introductory rates are $2 per million input tokens and $10 per million output tokens through November 30; rates rise to $4 and $20 on December 1, 2026.
A Close Score With a Lower Reported Cost
The MKA result draws attention because it pairs a near-leading score with a lower reported per-task cost. For organizations evaluating agents that must interpret visual or structured documents, that may make Carbon worth testing alongside higher-scoring systems. But the cost comparison is tied to this benchmark’s tasks and reporting method; it does not establish that Carbon will be cheaper for a company’s own workload.
The ranking also arrives as Carbon becomes more widely available. Before the November 10 rollout, it had been offered to a small group of cyber defenders through Google’s Fairwind Program, according to the source report. Broadening access gives paying users a chance to evaluate the model directly, while the score provides one published point of comparison with competitors.
For readers comparing model rankings, the main takeaway is not simply that Carbon came second. The three leading MKA scores are close enough to fall within the stated margin of error, and the benchmark was co-developed by Google DeepMind, whose model is among the entrants. Those details matter when interpreting both the ranking and the cost result.
Popular right now
Breville Barista Express Espresso Machine with Grinder, Black Truffle
Roborock Robot Vacuum and Mop with LiDAR Navigation and Self-Empty Dock
LG 55-Inch QNED73B Mini LED 4K Smart TV (2026)
As an affiliate, we earn on qualifying purchases.
Carbon’s Position Across Other Tests
Google describes Carbon as a higher-capability model built on Gemini 4 Argon for demanding work such as multi-step coding and debugging and open-ended scientific research. Its model card says Carbon uses more thinking tokens than Argon to explore possible answers and carry out extended verification. Google also says Argon remains the better fit for most software engineering and knowledge work, framing Carbon as a specialist rather than a universal replacement.
The model accepts text, images, audio and video and produces text. Its context window is up to one million tokens, with input and output counted together. Google lists low, medium and high thinking settings, with medium as the default. The model card also flags hallucinations, occasional slowness or timeouts, and higher token use, especially at higher thinking levels.
On Artificial Analysis’s Intelligence Index version 5.0.1, the source report gives Carbon a score of 60. Claude Fable 5.2 scores 65 at its maximum setting and 62 at high; Claude Opus 5.5 scores 59, and GPT-6 Astra 57. Artificial Analysis is also reported to find Carbon slow per task on its index evaluation: 25.6 minutes, compared with a 12.6-minute median for comparable models. These are separate evaluations from MKA and should not be combined into one ranking.
“Argon remains the ideal choice for the majority of software engineering and knowledge work.”
— Google, in the Gemini 4 Carbon model card
Limits of the MKA Comparison
The published material does not establish whether Carbon’s 2.3-point gap behind Fable 5.2 represents a reliable difference, since the source report says the top three results are within the margin of error. MKA’s task set is private, according to the source material, which means outside researchers cannot independently inspect the questions or reproduce the evaluation from the published tasks.
Google DeepMind co-developed the benchmark, and Carbon is one of the models it evaluates. That does not by itself invalidate the result, but it is relevant when weighing the test’s design and the reported advantage. The available details do not specify enough about the cost calculation to establish how closely the per-task amounts would match real-world use, where prompts, output length, retries and workload can vary.
It is also unclear how Carbon will perform on individual customers’ tasks, or how its speed, accuracy and costs compare under consistent settings outside these evaluations. Google’s model card discloses known issues, including occasional timeouts and high token use; the source material does not quantify how often those problems occur.
Access, Pricing and Further Testing
Developers and businesses can evaluate Carbon through Google’s paid developer and business offerings, while app and Search access remains restricted to Google AI Ultra subscribers. The announced API pricing changes on December 1, 2026, from introductory rates of $2 per million input tokens and $10 per million output tokens to $4 and $20, respectively.
The next useful evidence will come from evaluations that publish enough information for independent scrutiny and from customer testing on real tasks. Comparisons should account for model settings, response time, token use and the cost of completing a task, rather than relying on a single leaderboard position. No further MKA release or new benchmark milestone is specified in the source material.
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%. Claude Opus 5.5 scored 53.0%. The source report says the top three results are within the margin of error.
What does the MKA benchmark test?
MKA tests AI agents on 200 expert-built tasks across 10 sectors, including work involving blueprints, scans and charts rather than only plain text.
How much did Carbon cost per MKA task?
Google DeepMind’s report gives Carbon a cost of $5.63 per task, compared with $35.28 for Claude Fable 5.2. The figures refer to the benchmark evaluation and are not a universal price for customer workloads.
Who can use Gemini 4 Carbon?
Carbon is available through the paid Gemini API, Google AI Studio and Gemini for business. Access in the Gemini app and AI Mode in Search is limited to Google AI Ultra subscribers.
Does the second-place result prove Carbon is better for real-world work?
No. MKA is one benchmark, its tasks are private, and Google DeepMind co-developed it. The reported top scores are within the margin of error, and real-world performance and costs depend on the tasks and settings used.
Source: ThorstenMeyerAI.com
