🔍 Read the full analysis: The New MKA Test Ranks Gemini 4 Carbon Second on ThorstenMeyerAI.com
Get hardware and tech essentials delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
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 ranked it second behind Anthropic’s Claude Fable 5.2. Carbon scored 54.3% on MKA and cost $5.63 per task, compared with Fable 5.2’s 56.6% and $35.28, according to the benchmark report. The test was co-developed by Google DeepMind, and its private tasks cannot be independently inspected.
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 a benchmark that placed it second among three leading models. Carbon scored 54.3% on the new MKA test, behind Anthropic’s Claude Fable 5.2 at 56.6%, while the report put Carbon’s cost at $5.63 per task versus $35.28 for Fable 5.2.
MKA is designed to test AI agents on tasks involving blueprints, scans and charts, rather than only plain text. The benchmark comprises 200 expert-built tasks spanning 10 sectors. Claude Fable 5.2 scored 56.6%, Gemini 4 Carbon scored 54.3%, and Claude Opus 5.5 scored 53.0%. The source report says the three results fall within the margin of error, so the ranking does not establish a clear performance gap.
The cost figures come from Google DeepMind’s report on MKA: $5.63 per Carbon task and $35.28 per Fable 5.2 task. They are benchmark-specific figures, not a general estimate of what every real-world job will cost. The source does not give the full task-level pricing methodology, so the figures should be read as reported comparisons for this test.
Carbon is also ranked second on Artificial Analysis’s Intelligence Index version 5.0.1, according to the source material. It scores 60 at the high thinking setting; Fable 5.2 scores 62 at high and 65 at its maximum setting. These are separate results from MKA and should not be treated as interchangeable measures of model performance.
Gemini 4 Carbon ranks second on the new MKA test
Google opened Carbon to paying users on November 10, 2026, as a new benchmark compared three leading models on visual and document-heavy tasks.
How the models scored
All three results fall within the report’s margin of error.
Reported cost per MKA task
The benchmark report lists a lower task cost for Carbon in this evaluation.
- $5.63 per task
- $35.28 per task
What MKA tests
It evaluates agents on work that goes beyond plain text.
Visual materials
Tasks can involve blueprints, scans and charts.
Defined task set
The 200 expert-built tasks span 10 sectors.
Access and pricing
Availability depends on the Google product or channel.
Developers and businesses can use Carbon 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.
Listed rates are scheduled to rise on December 1, 2026, from $2 to $4 per million input tokens and from $10 to $20 per million output tokens.
What the MKA Ranking Shows
The result presents a trade-off for organizations considering advanced AI agents: Carbon scored close to the leading model on MKA while its reported cost per benchmark task was substantially lower. That may interest teams evaluating models for document-heavy work, where important information can be embedded in diagrams, scans or other visual material rather than in ordinary text.
But the ranking is not a universal verdict on which model is best. MKA covers a defined set of 200 tasks, and the reported scores are close enough that the source describes them as within the margin of error. Actual results and costs may differ with a company’s own data, task mix, settings and usage. The benchmark offers one comparison point, not a guarantee that Carbon will be cheaper or perform better in production.
Availability also comes with different access conditions. Carbon is offered through the paid Gemini API, Google AI Studio and Gemini for business. In the Gemini app and AI Mode in Search, the source says access is limited to Google AI Ultra subscribers. That distinction matters to users deciding whether the model is available through the product they already use.
Products Worth Considering
Apple MacBook Air 15-inch (2026) with M5 Chip, 16GB Memory, 1TB SSD, and Liquid Retina Display – Sky Blue
Apple MacBook Pro 16.3-inch with PROLED XDR Display, Silver
Logitech MX Mechanical Wireless Illuminated Keyboard, Clicky Switches
As an affiliate, we earn on qualifying purchases.
Carbon’s Launch and Access
Google describes Carbon as a top-tier model built on Gemini 4 Argon, intended for demanding tasks 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 choice for most software engineering and knowledge work, positioning Carbon as a more specialized option rather than a replacement for every task.
The model accepts text, images, audio and video, and returns text. Its context window is up to 1 million tokens, combining input and output, and its listed knowledge cutoff is March 2026, with a warning that knowledge may be less current in some areas. Google lists low, medium and high thinking settings, with medium as the default. The source says Carbon first went to a small group of cyber defenders through Google’s Fairwind Program around two weeks before its broader paid release; security features remain limited to verified defenders.
Google’s model card reports Carbon leading on 15 of 17 benchmarks selected by Google, but that comparison uses Google’s chosen tests and rival models and does not include Claude Fable 5.2, which the source says was released November 9. The same source notes that Carbon narrowly leads on some tests and loses to GPT-6 Astra on FrontierSWE 2 and Claude Opus 5.5 on OSWorld-2.1. Those results provide additional context, but they are distinct from the independent MKA ranking.
Limits of the MKA Evidence
MKA’s tasks are private, according to the source, which means outside researchers cannot inspect the items or independently reproduce the test from the task set. Google DeepMind co-developed the benchmark, and Carbon performed well on it. That does not by itself invalidate the results, but it makes independent scrutiny of the task design and scoring harder. The source material does not provide enough information to assess how the benchmark handles model prompting, repeated trials or variation between tasks.
The reported scores also come with a stated margin-of-error caveat: all three models are within that margin. The material does not specify the margin’s size or provide confidence intervals, so readers cannot tell from these figures alone how stable the ordering would be across repeated evaluations. It also remains unclear how closely MKA’s tasks reflect any particular company’s daily workload.
Pricing comparisons require care. The MKA report gives per-task costs, while Artificial Analysis reports costs per Intelligence Index task; those are different evaluations. Carbon’s listed API prices are $2 per million input tokens and $10 per million output tokens through November 30, 2026, rising to $4 and $20 from December 1, according to the source. Actual bills depend on token use, settings and task length. Google also warns of hallucinations, occasional timeouts and higher token consumption at higher thinking levels.
Pricing and Independent Checks
The next stated pricing milestone is December 1, 2026, when Carbon’s listed API rates are scheduled to rise from $2 per million input tokens and $10 per million output tokens to $4 and $20. Developers and businesses can evaluate the model through the paid channels identified by Google, while Gemini app and Search users need Google AI Ultra access, according to the source.
Further independent scrutiny of MKA would help clarify how meaningful the close scores are. The source material does not announce a date for public release of the task set, a repeat evaluation or a revised benchmark report. Until more detail is available, the published results support a narrow conclusion: Carbon placed second on MKA, with a lower reported cost per task than Fable 5.2, but the test does not establish a general performance or cost advantage across other uses.
Key Questions
What did Gemini 4 Carbon 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%, according to the source report. It says the three scores are within the margin of error.
What does the MKA benchmark test?
MKA uses 200 expert-built tasks across 10 sectors to test agents working with materials such as blueprints, scans and charts, rather than only plain text.
How much did each model cost per MKA task?
Google DeepMind’s report gives a cost of $5.63 per task for Carbon and $35.28 for Claude Fable 5.2. These are figures for the benchmark comparison, not a promise about costs for other workloads.
Who can use Gemini 4 Carbon?
According to the source, Carbon is available through the paid Gemini API, Google AI Studio and Gemini for business. In the Gemini app and AI Mode in Search, it is limited to Google AI Ultra subscribers.
Is the MKA ranking conclusive?
No. The source says the top three models are within the margin of error, and MKA’s tasks are private. Google DeepMind co-developed the test, so the reported ranking is useful as one comparison but cannot settle how models perform across all tasks.
Source: ThorstenMeyerAI.com
