🔍 Read the full analysis: Gemini 4 Carbon Comes In Second On MKA At A Lower Cost on ThorstenMeyerAI.com
Get the latest gadgets 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 it and Artificial Analysis introduced the MKA benchmark for agents working with visual documents. Carbon scored 54.3%, second to Claude Fable 5.2 at 56.6%; Google DeepMind reported a much lower MKA cost per task for Carbon, though the benchmark was co-developed by Google and its tasks are private.
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 for AI agents working with blueprints, scans and charts. Carbon placed second on MKA, scoring 54.3% against 56.6% for Anthropic’s Claude Fable 5.2, while Google DeepMind reported a substantially lower cost per task for Carbon: $5.63 versus $35.28.
MKA, short for Multimodal Knowledge and Agent benchmark, comprises 200 expert-built tasks across 10 sectors. Its tasks require models to extract and use information from visual and structured materials, rather than respond only to plain text. Claude Fable 5.2 scored 56.6%, Gemini 4 Carbon 54.3%, and Claude Opus 5.5 53.0%. The published report says the top three results fall within the margin of error, so the small score gaps should not be read as a definitive ranking of capability.
Google DeepMind’s report puts Carbon’s MKA cost at $5.63 per task, compared with $35.28 for Fable 5.2. That is a benchmark-specific cost estimate, not a general price for every use of either model. Carbon’s API price is temporarily $2 per million input tokens and $10 per million output tokens through November 30; Google says those rates rise to $4 and $20, respectively, on December 1.
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 describes Carbon as a model for demanding, multi-step work, including complex coding and scientific research, and says it uses more reasoning tokens than Gemini 4 Argon. The company also says Argon remains the better fit for most software engineering and knowledge-work tasks.
Gemini 4 Carbon Comes in Second at Lower MKA Cost
Google opened Carbon to paying customers as it introduced MKA, a benchmark for agents working with visual documents. Its near-leading score and lower reported task cost offer buyers a reason to test it on their own workflows.
How the scores compare
The top three results are within the report’s margin of error, so the small gaps may not show a meaningful capability difference.
| Model | MKA score | Reported cost per task |
|---|---|---|
| Claude Fable 5.2 | 56.6% | $35.28 |
| Gemini 4 Carbon | 54.3% | $5.63 |
| Claude Opus 5.5 | 53.0% | Not stated |
What MKA tests
The benchmark measures agents using visual and structured materials, beyond plain-text responses.
Blueprints and scans
Tasks ask models to extract information from visual documents and use it to complete work.
Charts and structured material
The benchmark covers visual information across 10 sectors.
Access and pricing
Availability depends on the product, and API rates are scheduled to change.
Paid Gemini API, Google AI Studio and Gemini for business are available; Gemini app and Search AI Mode access is limited to Google AI Ultra subscribers.
Temporary rates through November 30 are $2 per million input tokens and $10 per million output tokens. Google says they rise to $4 and $20 on December 1.
Measure accuracy, output length, retries, speed and total cost on your own materials.
A Lower-Cost Result on Visual Tasks
The MKA result gives buyers evidence about a category of work that conventional text benchmarks may not capture: agents interpreting documents, charts and technical imagery and using that information to complete tasks. Carbon’s near-leading score could make it worth testing for organizations that process such materials, particularly if the reported task cost holds in their own workflows.
The comparison is not a universal price-performance verdict. The quoted costs apply to the benchmark’s tasks and may not reflect a company’s prompt lengths, retries, integration costs or production volume. Artificial Analysis’ separate Intelligence Index also puts Carbon behind Fable 5.2: its November 11 figures show Carbon at 60, versus Fable 5.2 at 62 on the high setting and 65 at maximum. On that index, Carbon is cheaper than some competitors at standard pricing but not all.
Speed is another factor for deployment. Artificial Analysis describes Carbon as slow per task and verbose; its reported average is 25.6 minutes per index task, compared with a 12.6-minute median for comparable models. The right choice will depend on the task, latency needs and total operating cost—not MKA rank alone.
Products Worth Considering
Samsung Galaxy Tab A9+ 11-Inch Android Tablet, 64GB, Graphite
LG UltraGear 27-inch QHD OLED Gaming Monitor, 240Hz
Bose TV Speaker (2nd Gen) Compact Soundbar
As an affiliate, we earn on qualifying purchases.
From Restricted Preview to Paid Access
Carbon was first made available to a small group of cyber defenders through Google’s Fairwind Program about two weeks before the November 10 launch. Artificial Analysis lists October 28 as its release date. The model’s stronger security features remain limited to verified defenders, according to the source material; general paid availability does not mean those capabilities are open to all users.
Google’s model card reports Carbon leading on 15 of 17 benchmarks selected for its comparison with Gemini 4 Argon, GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1. Those figures are Google’s own scorecard, with results dated October 2026; the table does not include Claude Fable 5.2, released November 9. Carbon’s results vary by test: it leads on some long-context and automation measures but trails GPT-6 Astra on FrontierSWE 2 and Claude Opus 5.5 on OSWorld-2.1.
The independent Intelligence Index provides a different snapshot. Its current version, v5.0.1, scores Carbon (high) at 60; the older v5.0 figure shown in Google’s model card had Carbon at 59 and Claude Opus 5.5 at 58. Because models, settings and benchmark versions differ, results from the Google scorecard, MKA and the index are not interchangeable.
““tasks that test the limits of a model’s intelligence””
— Google DeepMind, in the Gemini 4 Carbon model card
Benchmark Limits and Real-World Costs
MKA was launched by Google DeepMind and Artificial Analysis, and Google’s model is among those evaluated. The report’s tasks are private, so outside researchers cannot inspect them or independently reproduce the full test set. The reported margin of error also means the ordering of the top three may not indicate a meaningful performance gap. These points limit how far the result can be generalized from the published score.
The $5.63 and $35.28 figures are reported benchmark costs, but the available material does not fully specify how those costs would translate to each customer’s workload. It also does not establish whether Carbon’s lower MKA cost persists across different settings or task types. Google lists hallucinations, occasional slowness or timeouts, and higher token use—particularly at higher thinking levels—as known limitations. Its context window can reach one million tokens across input and output, and its stated knowledge cutoff is March 2026, with some domains potentially less current.
It remains unclear how Carbon will perform in broad commercial use, how many users will receive access under each plan, and whether the MKA results will be replicated by independent tests with public tasks. The benchmark score alone cannot settle those questions.
Access, Pricing and Independent Tests
Developers and businesses can test Carbon through Google’s paid developer and business offerings, while app and Search access is restricted to Ultra subscribers. Buyers considering the model should account for the scheduled December 1 price increase and test it on their own documents and workflows, including the costs of longer outputs, retries and delays.
The next useful evidence will come from broader independent evaluations and real-world reports that compare task accuracy, speed and total cost under clearly described conditions. Until then, MKA offers an initial result—not a final measure of which model is best for every visual-document or agent task.
Key Questions
What did Gemini 4 Carbon score on MKA?
Carbon scored 54.3%, placing second behind Claude Fable 5.2 at 56.6%. Claude Opus 5.5 scored 53.0%. The report says the top three are within the margin of error.
How much cheaper was Carbon on the benchmark?
Google DeepMind reported a cost of $5.63 per MKA task for Carbon and $35.28 for Claude Fable 5.2. These are reported benchmark costs, not guaranteed prices for other workloads.
Who can use Gemini 4 Carbon?
It is available through the paid Gemini API, Google AI Studio and Gemini for business. In the Gemini app and AI Mode in Search, Google limits access to AI Ultra subscribers.
What does MKA test?
MKA contains 200 tasks across 10 sectors designed to test AI agents working with materials such as blueprints, scans and charts, rather than text alone.
Can the MKA ranking be independently checked?
Not fully from the published materials: the MKA tasks are private, and Google DeepMind helped develop the benchmark. The reported results provide a comparison, but outsiders cannot inspect and rerun the complete task set.
Source: ThorstenMeyerAI.com
