🔍 Read the full analysis: A Second-Place MKA Finish At A Fraction Of The Leader’s Cost on ThorstenMeyerAI.com
Get smart everyday buys 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 customers on November 10, the same day Google DeepMind and Artificial Analysis introduced the MKA benchmark for AI agents working with visual and structured information. Carbon placed second with 54.3%, compared with Claude Fable 5.2’s 56.6%; Google DeepMind reported a per-task cost of $5.63 for Carbon versus $35.28 for Fable. The benchmark was co-developed by Google DeepMind, its tasks are private, and the top scores are within the stated margin of error.
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 launched a benchmark for AI agents handling blueprints, scans and charts. Carbon scored 54.3% on MKA, placing second behind Anthropic’s Claude Fable 5.2 at 56.6%; Google DeepMind’s report puts Carbon’s cost at $5.63 per task, compared with $35.28 for Fable.
MKA, short for Multimodal Knowledge Assessment, comprises 200 expert-built tasks across 10 sectors. It tests agents on material such as diagrams and other visual or structured information, rather than relying only on plain-text prompts. Claude Opus 5.5 scored 53.0%, behind Carbon. The source report says the leading results are within the margin of error, so the small gap between Fable and Carbon should not be treated as a definitive measure of superiority.
The task-cost figures come from Google DeepMind’s MKA report. They indicate a large cost difference in that benchmark’s evaluation, not a general price guarantee across workloads or a measure of every expense involved in deploying a model. The supplied source does not give a detailed accounting of how the per-task costs were calculated.
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. Google set API pricing at $2 per million input tokens and $10 per million output tokens through November 30, 2026. From December 1, the listed rates rise to $4 and $20, respectively.
Gemini 4 Carbon Places Second At a Lower Reported Cost
Carbon scored close to Claude Fable 5.2 on the new MKA benchmark, while Google DeepMind reported a much lower cost per task. The ranking is within the stated margin of error.
What MKA Measures
The benchmark tests agents on visual and structured information, including diagrams, scans and charts.
MKA includes 200 tasks across 10 sectors.
Agents work with more than plain-text prompts.
Outside researchers cannot inspect or reproduce the full evaluation from its tasks.
The Cost Gap in Context
The reported figures compare cost per task in this benchmark evaluation.
- $5.63 reported per MKA task
- 54.3% score, close to the leader
- $35.28 reported per MKA task
- 56.6% score, within the stated margin of error
Availability and API Pricing
Developers and businesses can access Carbon through paid Google products; app and Search access requires Google AI Ultra.
| API rate | Through November 30 | From December 1 |
|---|---|---|
| Input, per million tokens | $2 | $4 |
| Output, per million tokens | $10 | $20 |
What the Cost Gap Shows
The result puts performance and cost side by side for buyers considering high-end AI agents. In this evaluation, Carbon came within 2.3 percentage points of Fable’s score while its reported task cost was much lower. For companies running repeated tasks on complex documents or visual information, cost per task may affect which model is practical to test or deploy.
That comparison is limited to the MKA setup and its reported figures. A lower benchmark cost does not establish that Carbon will be cheaper for every customer: actual spending can vary with task length, model settings, input volume and the amount of reasoning used. Google also says Carbon uses more thinking tokens than its Gemini 4 Argon model, and Artificial Analysis describes it as slow and verbose on its own index tasks.
For users, availability is a separate consideration from benchmark rank. Developers and businesses can access Carbon through paid products, while app and Search access requires an AI Ultra subscription. Google’s published token rates also change on December 1, making the date relevant for teams estimating API expenses.
Popular right now
Shark Rotator Lift-Away Upright Vacuum with DuoClean PowerFins, Wine Purple, LA702
Ninja Crispi Microwave and Air Fryer with FusionCrisp, Silver, MW101
Shark ZU102 Rotator Pet Upright Vacuum with PowerFins HairPro & Odor Neutralizer Technology, Charcoal
As an affiliate, we earn on qualifying purchases.
Carbon’s Release and Benchmark Record
Carbon had previously been offered to a small group of cyber defenders through Google’s Fairwind Program. The source says Artificial Analysis lists October 28 as its release date; Google’s broader paid launch followed on November 10. Security features associated with that program remain limited to verified defenders, according to the supplied material.
Google describes Carbon as a top-tier model built on Gemini 4 Argon, intended for demanding work such as multi-step coding, debugging and scientific research. Its model card lists text, image, audio and video inputs, text output, and a context window of up to one million tokens across input and output. It offers low, medium and high thinking settings, with medium as the default. Google says Argon remains the better fit for most software engineering and knowledge work.
The MKA result is not the only published comparison. On Artificial Analysis’ Intelligence Index v5.0.1, Carbon at high reasoning scored 60. Fable 5.2 scored 65 at its maximum setting and 62 at high. Those results use a different evaluation from MKA and should not be combined with its percentages. Google’s model card reports Carbon leading 15 of 17 benchmarks it selected, but that comparison was based on October results and did not include Fable 5.2, released November 9, according to the source.
““tasks that test the limits of a model’s intelligence””
— Google DeepMind, describing Gemini 4 Carbon in its model card
Limits of the MKA Comparison
Google DeepMind co-built MKA, and its own model is among the top scorers. That does not by itself invalidate the results, but it is relevant when evaluating the benchmark’s independence. The source says the tasks remain private, meaning outside researchers cannot inspect the full set or independently reproduce the evaluation from those tasks.
The top three scores are described as within the margin of error, but the supplied material does not state the exact uncertainty range or provide enough detail to assess the statistical separation between the models. It is therefore unclear whether the small ranking differences reflect meaningful performance gaps.
The reported costs also need qualification. The source attributes the $5.63 and $35.28 figures to Google DeepMind’s report but does not explain the full cost methodology. Carbon’s speed and verbosity on Artificial Analysis’ separate index are additional considerations, but those measurements cannot be directly converted into MKA completion times. Google’s model card also lists hallucinations, occasional slowness or timeouts, and increased token use at higher thinking levels as limitations.
Pricing and Independent Testing
The immediate change for API customers is the scheduled December 1 price increase: Google’s listed rates move from $2 per million input tokens and $10 per million output tokens to $4 and $20. Buyers comparing models will need to assess their own workloads, since benchmark task costs and token prices describe different measures of expense.
Further independent evaluations could help clarify how Carbon performs on tasks beyond the private MKA set and whether the reported cost advantage holds under other testing conditions. The available source material does not identify a date for additional MKA results or say whether the benchmark’s tasks will later be made public. Until more detail is available, the confirmed picture is a close second-place MKA score, a lower task-cost figure reported by Google DeepMind, and unresolved questions about reproducibility and the size of the performance gap.
Key Questions
What was Gemini 4 Carbon’s score on MKA?
Carbon scored 54.3% on the new benchmark. Claude Fable 5.2 scored 56.6%, and Claude Opus 5.5 scored 53.0%. The source says the top three are within the margin of error.
How much did each model cost per MKA task?
Google DeepMind’s report lists $5.63 per task for Carbon and $35.28 for Claude Fable 5.2. These are reported benchmark costs, not guaranteed prices for customer workloads.
Who can use Gemini 4 Carbon now?
It 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, according to the source.
When do Carbon’s listed API prices change?
Google’s listed rates are $2 per million input tokens and $10 per million output tokens through November 30, 2026. From December 1, 2026, the rates are scheduled to become $4 and $20.
How reliable is the MKA ranking?
The benchmark was launched by Google DeepMind and Artificial Analysis, and the source says its tasks are private. The reported scores put the leading models within the margin of error, but the supplied material does not give the exact uncertainty range or enough information to reproduce the test independently.
Source: ThorstenMeyerAI.com
