🔍 Read the full analysis: How Gemini 4 Carbon Delivers A Second-Place MKA Result For Less 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 users on November 10, the same day it and Artificial Analysis introduced the MKA benchmark for agents working with visual and structured materials. Carbon scored 54.3%, placing second behind Claude Fable 5.2’s 56.6%; Google’s report puts Carbon’s cost at $5.63 per MKA task versus $35.28 for Fable. The benchmark was co-developed by Google and its tasks are private, limiting independent scrutiny.
MKA tests AI agents on **200 expert-built tasks across 10 sectors**, using materials such as blueprints, scans and charts rather than relying only on plain text. In the published results, **Claude Fable 5.2 scored 56.6%**, Gemini 4 Carbon 54.3% and Claude Opus 5.5 53.0%. The source report says the top three results are **within the margin of error**, so the small score gaps should not be treated as definitive evidence of a clear performance separation.
Google DeepMind’s report gives Carbon a cost of **$5.63 per MKA task** and Fable 5.2 a cost of **$35.28**. Those figures make the benchmark’s price comparison a central part of the result, but they are reported estimates tied to this test—not a guarantee of what a customer will pay for every real-world task. The available material does not provide enough detail to independently verify the task-level cost calculation.
Carbon is Google’s **top-tier model**, built on Gemini 4 Argon and designed for demanding work such as complex coding, debugging and scientific research. Google says it uses more thinking tokens than Argon. Its model card lists support for text, images, audio and video inputs, text output, and a context window of up to **1 million tokens**. Google also cites hallucinations, occasional slowdowns or timeouts, and higher token use as known limitations.
Gemini 4 Carbon: near the top, at lower reported cost
Google’s model placed second on the new MKA benchmark. Its reported task cost was lower than the leader’s, but private tasks and benchmark uncertainty limit the comparison.
A close result, with a cost gap
The top three scores are reported within the margin of error.
| Model | MKA score | Reported task cost |
|---|---|---|
| Claude Fable 5.2 | 56.6% | $35.28 |
| Gemini 4 Carbon | 54.3% | $5.63 |
| Claude Opus 5.5 | 53.0% | Not provided |
What MKA tests
The benchmark challenges agents to work with visual and structured materials, not only plain text.
Real-world inputs
Tasks use materials such as blueprints, scans and charts.
Ten sectors
Expert-built tasks span a defined set of work, not every customer workflow.
Access and pricing
Paid access opened November 10, 2026; introductory API rates are scheduled to change after November 30.
Developers and businesses can use Carbon through the Gemini API and Google AI Studio.
$2 per million input tokens and $10 per million output tokens through November 30.
The stated rates become $4 per million input tokens and $20 per million output tokens.
A Lower-Cost Result With Caveats
The result offers companies a concrete reason to compare **quality and cost together** when choosing an AI agent. On MKA, Carbon finished close to the leader by score while Google’s report assigned it a substantially lower per-task cost. For organizations processing documents, diagrams or other visual material, that combination could make Carbon worth testing, particularly when tasks are numerous and costs accumulate.
But the benchmark does not establish that Carbon is the best choice across all work. MKA covers a defined set of **200 tasks**, and the score differences among the leading models fall within the reported margin of error. Google and Artificial Analysis developed the test together, and the task set is private, leaving outside researchers unable to inspect the questions or reproduce the evaluation from the task materials.
Availability also comes with pricing conditions. Google’s introductory API prices are **$2 per million input tokens and $10 per million output tokens through November 30, 2026**; the stated rates rise to $4 and $20 from December 1. Carbon’s slower performance on another published comparison may also matter to customers: Artificial Analysis describes it as slow and verbose on its Intelligence Index tests. A low benchmark task cost does not, on its own, capture every trade-off in latency, reliability or product fit.
Popular right now
De'Longhi Rivelia Automatic Espresso Machine with Grinder and Milk Frother, Beige
Ninja Crispi Microwave & Air Fryer with FusionCrisp Technology, 14-in-1, Black (MW101BK1)
Bose QuietComfort True Wireless Earbuds with Active Noise Cancellation, Bluetooth 5.3, USB-C Charging, IPX4, Black
As an affiliate, we earn on qualifying purchases.
From Restricted Preview to Paid Access
Carbon was previously available to a small group of cyber defenders through Google’s **Fairwind Program**. Artificial Analysis lists an October 28 release date. Google’s November 10 announcement expanded access to paying developers and businesses through the Gemini API and Google AI Studio, as well as Google AI Ultra subscribers in the Gemini app and AI Mode in Search. Its stronger security features remain limited to verified defenders, according to the supplied source.
Google’s model card reports top results on **15 of 17 benchmarks** it selected, against four models named in that comparison. The card’s results are dated October 2026 and do not include Claude Fable 5.2, which the source says was released November 9. The card reports narrow leads on some tests, including SWE-sweep, and losses to GPT-6 Astra on FrontierSWE 2 and Claude Opus 5.5 on OSWorld-2.1. These results are useful context, but they are not the same evaluation as MKA.
On Artificial Analysis’s separate **Intelligence Index v5.0.1**, Carbon at its high setting scores 60. Fable 5.2 scores 62 at high and 65 at its maximum setting. The index combines nine tests and is distinct from MKA; its rankings and cost estimates should not be mixed with the MKA percentages and per-task figures.
““tasks that test the limits of a model’s intelligence””
— Google, in the Gemini 4 Carbon model card
Benchmark Limits and Open Questions
The MKA results do not settle whether Carbon is meaningfully worse than Fable 5.2: the source says the top three scores are **within the margin of error**, but it does not provide the underlying uncertainty estimates or enough detail here to assess statistical significance. The tasks are **private**, so independent researchers cannot directly inspect them or rerun the same test.
Google’s role as a co-developer of MKA is another reason to treat the results as a benchmark finding rather than a neutral, final verdict. The available source does not explain how task selection, scoring and cost accounting were independently audited. It also does not establish whether the reported per-task costs reflect typical customer workloads, or how those estimates change with task complexity and usage settings.
It remains unclear how Carbon performs on customers’ own materials and workflows, including the balance between accuracy, latency and total spending. Google lists occasional timeouts and hallucinations among the model’s limitations, but the supplied material does not quantify their frequency. The benchmark score alone cannot answer those operational questions.
Pricing Changes and Real-World Tests
Google’s stated introductory API rates run through **November 30, 2026**, with higher prices scheduled from December 1. Developers and businesses can access Carbon through the paid Gemini API and Google AI Studio; access in the consumer Gemini app and Search’s AI Mode is limited to Google AI Ultra subscribers, according to the source.
The next useful evidence will come from independent evaluations that test Carbon on reproducible tasks and from customers reporting performance on their own work. Comparisons should identify the model settings, task mix, pricing assumptions and time required, rather than relying on a single score. No further benchmark release or independent audit is confirmed in the supplied material.
Key Questions
What did Gemini 4 Carbon score on MKA?
Carbon scored **54.3%** on the new MKA 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 does Carbon cost per MKA task?
Google DeepMind’s report puts Carbon at **$5.63 per task**, compared with $35.28 for Fable 5.2. These are reported benchmark costs, not necessarily the price of every customer workload.
Who can use Gemini 4 Carbon?
Paying developers and businesses can use it through the Gemini API and Google AI Studio. In the Gemini app and AI Mode in Search, it is available to **Google AI Ultra subscribers**, according to the supplied source.
Why should the MKA ranking be treated cautiously?
Google co-developed MKA with Artificial Analysis, and the benchmark’s **200 tasks are private**. The supplied report also says the leading scores are within the margin of error, limiting what can be concluded from their ranking.
When does Carbon’s introductory pricing end?
The supplied source says API prices of **$2 per million input tokens and $10 per million output tokens** apply through November 30, 2026. It lists rates of $4 and $20, respectively, from December 1.
Source: ThorstenMeyerAI.com
