🔍 Read the full analysis: Gemini 4 Carbon Lands Second On The New MKA Test on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Google opened Gemini 4 Carbon to paying developers, businesses and Google AI Ultra subscribers on November 10, the same day a new benchmark for agents handling visual documents was introduced. Carbon scored 54.3% on MKA, second to Claude Fable 5.2 at 56.6%; Google DeepMind reported a lower per-task cost for Carbon, while the benchmark’s private tasks and Google’s role in its development limit independent scrutiny.
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 introduced the MKA benchmark for AI agents working with materials such as blueprints, scans and charts. Carbon scored 54.3%, placing second behind Anthropic’s Claude Fable 5.2 at 56.6%; Google DeepMind’s report put Carbon’s cost at $5.63 per MKA task, compared with $35.28 for Fable 5.2.
The MKA test comprises 200 expert-built tasks across 10 sectors. Claude Fable 5.2 led the published results, with Carbon next at 54.3% and Claude Opus 5.5 at 53.0%. The source report says the scores for the top three are within the margin of error, so the ranking does not establish a clear performance gap. Google DeepMind reported the per-task cost figures; they are tied to this benchmark and should not be read as general prices for every workload.
Carbon is Google’s top-tier Gemini 4 model, built on Gemini 4 Argon and intended for difficult, multi-step tasks such as coding and scientific research. Its model card lists text, image, audio and video inputs, text output, and a context window of up to 1 million tokens combined across input and output. Google says Carbon uses more thinking tokens than Argon to examine possible solutions and verify results, and describes Argon as the better fit for most software engineering and knowledge work.
Access depends on the product. 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. The listed API price is $2 per million input tokens and $10 per million output tokens through November 30, then $4 and $20 respectively from December 1, 2026. Google lists hallucinations, occasional slowness or timeouts, and higher token use—especially at higher thinking settings—as limitations.
Gemini 4 Carbon Lands Second On The New MKA Test
Google’s model ranked near the top on a new test for agents handling visual documents. Its reported cost was lower, but the benchmark’s private tasks limit independent scrutiny.
How the top results compare
The reported scores for the top three are within the margin of error, so the ranking does not establish a clear performance gap.
Reported cost per MKA task
Google DeepMind reported these benchmark-specific estimates; they may not predict costs on other workloads.
- $5.63 per task
- $35.28 per task
Who can access Carbon
Availability depends on the Google product.
| Product | Access |
|---|---|
| Gemini API, AI Studio, Gemini for business | Paid users |
| Gemini app and AI Mode in Search | Google AI Ultra subscribers |
What to weigh before choosing
Carbon is designed for difficult, multi-step tasks, with practical trade-offs for teams and users.
Visual and structured work
MKA tests agents on materials such as blueprints, scans and charts.
Private benchmark tasks
Outsiders cannot inspect the tasks or independently reproduce the results.
Time and token use
Google lists hallucinations, occasional slowness or timeouts, and higher token use as limitations.
API price change
Google’s listed rates change on December 1, 2026.
$2 per million input tokens and $10 per million output tokens.
$4 per million input tokens and $20 per million output tokens.
Cost and Scores for Agent Work
The release gives organizations a new option for agent tasks involving visual or structured information, not just text. MKA’s tasks are designed around materials such as scans, charts and blueprints, making its results relevant to companies considering models for document-heavy workflows. Carbon’s second-place result suggests it is competitive on this particular test, while the reported cost difference could matter to teams running many tasks.
Those figures do not settle which model is best for a company. The benchmark’s top three scores fall within the reported margin of error, and cost per task on MKA may not predict cost for another workload. Separately, Artificial Analysis’ Intelligence Index v5.0.1 gives Carbon (high) a score of 60, below Fable 5.2 at 62 in its high setting and 65 at its maximum setting. That broader index and MKA measure different things, so their rankings should not be combined into a single verdict.
For users, availability is also not universal: app and Search access requires Google AI Ultra, while developers and businesses can use paid Google services. Google says Carbon is designed for harder tasks and can use more tokens and time, leaving a practical trade-off between capability, response speed and expense.
Products Worth Considering
Logitech MX Mechanical Wireless Illuminated Keyboard, Tactile Quiet, Graphite
LG 27US500-W 27-Inch 4K UHD IPS Monitor
Vari Ergo Electric Standing Desk, 54×26, Light Wood/White
As an affiliate, we earn on qualifying purchases.
From Limited Access to Paid Release
Carbon had previously been made available to a small group of cyber defenders through Google’s Fairwind Program. The source report says Artificial Analysis lists October 28 as its release date; Google opened broader paid access on November 10. The model’s security features remain restricted to verified defenders, according to the report, so the general release does not mean all security-related capabilities are available to every customer.
Google’s model card reports leading results for Carbon on 15 of 17 benchmarks it selected, but those comparisons are not a complete independent ranking. The card’s comparison set includes Gemini 4 Argon, GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1; it does not include Fable 5.2, which the source says was released November 9. The card also dates its reported results to October 2026. On two listed tests Carbon narrowly trails competitors: GPT-6 Astra leads on FrontierSWE 2, and Claude Opus 5.5 leads on OSWorld-2.1.
In the current Artificial Analysis Intelligence Index version cited by the source, Carbon scores 60, while Claude Fable 5.2 scores 65 at maximum and 62 at high. Artificial Analysis also describes Carbon as slow and verbose on that evaluation: 25.6 minutes per index task, compared with a 12.6-minute median for comparable models. That result is specific to the index’s tasks, not a measure of every real-world response.
““tasks that test the limits of a model’s intelligence””
— Google, in the Gemini 4 Carbon model card
Limits of the MKA Results
MKA’s task set is private, according to the source report, preventing outsiders from checking the questions and independently reproducing the results. Google DeepMind helped build the benchmark, and Google’s model placed near the top. That does not by itself invalidate the scores, but it is relevant when evaluating how much weight to give them. The available source does not provide enough detail to independently assess the benchmark’s full scoring procedure or the stated margin of error.
It is also unclear how closely MKA performance will translate to commercial work. The published results cover 200 tasks across 10 sectors, but do not establish how Carbon compares on every company’s documents, tools or operating conditions. The reported $5.63 and $35.28 costs are MKA task estimates from Google DeepMind’s report, not a guarantee of costs on other tasks. Google’s model card separately flags hallucinations, latency and token use as limitations.
The broader rankings are not directly interchangeable: Google’s model card uses tests selected by Google and results dated October, MKA is a new benchmark co-developed by Google DeepMind, and Artificial Analysis combines a separate set of evaluations. Carbon’s lead on some Google-selected tests, second place on MKA and second place on the cited index are distinct findings, not one unified ranking.
Price Changes and Independent Checks
The next published pricing change is scheduled for December 1, 2026, when Google’s listed API rates rise from $2 per million input tokens and $10 per million output tokens to $4 and $20. Developers evaluating Carbon can compare its performance and costs on their own workloads, while noting Google’s warnings about longer processing times, timeouts and higher token consumption at stronger thinking settings.
Further confidence in MKA’s results will depend on how the benchmark is documented and whether independent evaluations can test comparable tasks. The source material does not identify a specific date for additional MKA results or an independent release of its private tasks. Google’s model card and the Artificial Analysis index provide other comparison points, but their differing test sets mean readers should look for further results rather than treat one score as decisive.
Key Questions
What was Gemini 4 Carbon’s score on the MKA benchmark?
Carbon scored 54.3% on MKA, behind Claude Fable 5.2 at 56.6% and ahead of Claude Opus 5.5 at 53.0%. The source report says the top three results are within the margin of error.
What does the MKA benchmark test?
MKA contains 200 expert-built tasks across 10 sectors for AI agents handling materials such as blueprints, scans and charts, rather than only plain text.
Who can use Gemini 4 Carbon?
Google made Carbon 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.
How much does Gemini 4 Carbon cost?
Google’s listed API price is $2 per million input tokens and $10 per million output tokens through November 30, 2026. From December 1, the rates are scheduled to become $4 and $20. These API rates are separate from the MKA per-task cost estimates.
Can the MKA ranking be independently verified?
The source report says MKA’s tasks remain private, limiting outside reproduction of the results. Google DeepMind helped develop the benchmark, so the result is useful as a published comparison but has limitations for independent verification.
Source: ThorstenMeyerAI.com
