🔍 Read the full analysis: Second Place On The New MKA Test: Gemini 4 Carbon’s Cost Advantage 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 users on November 10, the same day Google DeepMind and Artificial Analysis introduced the MKA benchmark for agents working with visual and structured documents. Carbon scored 54.3%, narrowly behind Claude Fable 5.2 at 56.6%, while Google’s report put Carbon’s cost at $5.63 per task versus $35.28 for Fable. The benchmark’s tasks are private, Google helped build it, and the top three 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 documents such as blueprints, scans and charts. Carbon ranked second at 54.3%, behind Anthropic’s Claude Fable 5.2 at 56.6%, but Google’s report put Carbon’s cost at $5.63 per task, compared with $35.28 for Fable.
The new MKA benchmark comprises 200 expert-built tasks across 10 sectors. It tests agents on information presented in varied formats rather than only plain text. Claude Fable 5.2 scored 56.6%, Gemini 4 Carbon 54.3%, and Claude Opus 5.5 53.0%. The report says the scores of the top three are within the margin of error, so the ranking does not establish a clear performance gap.
Google DeepMind’s report lists Carbon’s MKA cost at $5.63 per task and Fable 5.2’s at $35.28. Those are benchmark-specific costs, not a universal estimate of what a company would pay for its own workload. The source material does not detail the cost calculation or provide results for every model configuration, so the figures should be read as reported comparisons for this test.
Carbon is available through the paid Gemini API and Google AI Studio, and to businesses using Gemini. In the Gemini app and AI Mode in Search, access is limited to Google AI Ultra subscribers. Google set introductory API pricing through November 30 at $2 per million input tokens and $10 per million output tokens; from December 1, the listed rates are $4 and $20. The product is a higher-compute model designed for difficult multi-step work, and Google says Gemini 4 Argon remains the better fit for most software engineering and knowledge work.
Gemini 4 Carbon places second with a reported cost edge
Carbon scored close to the top on a new document focused agent benchmark, while Google reported a much lower cost per task than the highest scoring model.
How the models scored
MKA tests agents on visual and structured documents across 10 sectors.
What the cost comparison means
Google’s figures are specific to this benchmark and are not a forecast for a company’s own workload.
- Lower reported task cost may matter for repeated work.
- Carbon handles documents such as charts, scans and blueprints.
- Latency and token use can affect real world costs.
- Artificial Analysis describes Carbon as slow per task.
Access and API pricing
Availability varies by product, and introductory rates are time limited.
Cost and Performance on MKA
The results put a practical trade-off in focus for teams choosing an AI agent: Carbon scored slightly below Fable on MKA but had a much lower reported cost per task. If that cost difference holds on a company’s own tasks, it could matter for repeated or large-scale work. The benchmark alone, however, does not show whether the models perform equally well on any particular business process.
The introduction of MKA also adds a new comparison point for models expected to interpret visual and structured materials. Such work can include extracting information from diagrams or charts, where a text-only score may not reflect an agent’s usefulness. But the available results do not establish that MKA predicts performance across all sectors or real-world deployments.
Speed may also affect the cost calculation in practice. Artificial Analysis describes Carbon as slow per task and verbose, reporting 25.6 minutes per task on its separate Intelligence Index tests, against a 12.6-minute median for comparable models. That is not an MKA timing result, and it should not be treated as one; it does signal that users may need to weigh latency and token use alongside the benchmark’s cost figure.
Popular right now
KitchenAid Artisan 5-Quart Tilt-Head Stand Mixer, Blue Velvet
LG 77-Inch OLED evo C6 4K Smart TV
Ninja Crispi Microwave & Air Fryer with FusionCrisp Technology, 14-in-1, Black (MW101BK1)
As an affiliate, we earn on qualifying purchases.
From Limited Access to Paid Release
Carbon was first made available to a small group of cyber defenders through Google’s Fairwind Program about two weeks before the broader paid release. Google says the model’s stronger security features remain limited to verified defenders. Its model card describes Carbon as a model for demanding tasks, including complex coding and open-ended scientific research, that can use more reasoning tokens than Gemini 4 Argon.
Google’s model card reports that Carbon led on 15 of 17 benchmarks selected for its comparison, against four listed rivals. That is a company-selected scorecard, and it did not include Claude Fable 5.2, which the source says was released November 9. Google dates the results in its card to October 2026. On the separate Artificial Analysis Intelligence Index version 5.0.1, Carbon at its high setting scored 60; Fable scored 62 at high and 65 at its maximum setting. These are distinct evaluations and should not be combined with MKA results as if they measured the same capabilities.
The model accepts text, images, audio and video, produces text, and has a context window of up to one million tokens, counting input and output together. Google lists low, medium and high thinking levels, with medium as the default. Its model card also warns of hallucinations, occasional slowness or timeouts, and higher token use, especially at higher thinking settings.
““tasks that test the limits of a model’s intelligence””
— Google, in the Gemini 4 Carbon model card
Limits of the MKA Comparison
MKA’s task set is private, so outside researchers cannot inspect the questions or independently reproduce the reported results from the task materials. Google DeepMind helped build the benchmark, and its own model ranks near the top. That does not by itself invalidate the results, but it makes independent replication and disclosure of the test design important for judging how broadly the scores apply.
The source says the leading scores fall within the margin of error but does not give the margin’s size or further detail on statistical methods. It is also not clear from the material how the per-task cost was calculated, which model settings were used for each cost figure, or how much those estimates would vary with different prompts and workloads. The reported comparison therefore supports a cautious reading, not a guarantee that Carbon will cost one-sixth as much in production.
Google’s access terms also differ by product, and the introductory API prices expire on November 30. The supplied information does not specify whether all regional availability, rate limits or account requirements are the same. Those details could affect who can use Carbon and at what cost.
Pricing Changes December 1
The next confirmed pricing date is December 1, 2026, when Google’s listed API rates are set to rise from $2 to $4 per million input tokens and from $10 to $20 per million output tokens. Developers considering the model can compare it with alternatives using their own tasks, tracking not just accuracy but latency, token consumption and the cost of reviewing errors.
Further independent testing could clarify whether MKA’s ranking holds when the benchmark is replicated or when models are evaluated on disclosed tasks. For now, the published evidence establishes Carbon’s second-place MKA score and Google’s reported cost comparison; it does not settle how the model will perform across production workloads.
Key Questions
What was Gemini 4 Carbon’s score on MKA?
Carbon scored 54.3%. Claude Fable 5.2 scored 56.6%, and Claude Opus 5.5 scored 53.0%. The report says the top three are within the margin of error.
How much did Carbon cost per MKA task?
Google DeepMind’s report lists Carbon at $5.63 per task, compared with $35.28 for Claude Fable 5.2. The source does not provide enough detail to treat those benchmark costs as a forecast for other workloads.
Who can use Gemini 4 Carbon?
It is available to paying developers through the Gemini API and Google AI Studio, and to businesses. In the Gemini app and AI Mode in Search, the source says access is limited to Google AI Ultra subscribers.
Is Carbon clearly better than the other models?
The MKA results do not establish a clear lead: Carbon placed second, and the report says the top three scores are within the margin of error. Results also vary across other benchmarks and model settings.
What is not known about the MKA results?
The benchmark tasks are not public, limiting outside verification. The supplied material also does not specify the margin-of-error size or enough detail about how per-task costs were calculated.
Source: ThorstenMeyerAI.com
