🔍 Read the full analysis: Gemini 4 Carbon: MKA Runner-Up At One-Sixth Of The 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 developers, businesses and Google AI Ultra subscribers on November 10, 2026. On the new MKA benchmark, it scored 54.3%, behind Claude Fable 5.2’s 56.6%, while Google DeepMind reported a per-task cost of $5.63 versus $35.28 for Fable. The benchmark was co-developed by Google DeepMind and Artificial Analysis, and its tasks are private, limiting outside verification.
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 published results from their new MKA benchmark. Carbon scored 54.3%, second to Anthropic’s Claude Fable 5.2 at 56.6%, while Google DeepMind reported that Carbon cost $5.63 per MKA task compared with $35.28 for Fable.
MKA tests AI agents on 200 expert-built tasks across 10 sectors. Unlike text-only evaluations, the tasks require models to interpret material such as blueprints, scans and charts. Claude Fable 5.2 placed first, Carbon second and Claude Opus 5.5 third at 53.0%. The three results are within the benchmark’s stated margin of error, so the ranking does not establish a clear performance gap among the top models.
Google’s published access terms vary by product. Carbon is available through the paid Gemini API, Google AI Studio and Gemini for business; in the Gemini app and AI Mode in Search, access is limited to Google AI Ultra subscribers. Google lists introductory API pricing of $2 per million input tokens and $10 per million output tokens through November 30. From December 1, those rates rise to $4 and $20 per million tokens.
Google’s model card reports that Carbon leads on 15 of 17 benchmarks selected by Google, but that scorecard uses results dated October 2026 and does not include Claude Fable 5.2, released November 9. In Artificial Analysis’ newer 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. These evaluations use different tests and should not be treated as interchangeable rankings.
Gemini 4 Carbon: MKA Runner-Up At One-Sixth Of The Cost
Google opened Carbon to paying users as it placed second on a new visual-agent benchmark. Its reported task cost was far lower than the leader’s, but private tasks and a stated margin of error limit what the ranking proves.
What MKA tests
The benchmark evaluates agents working with visual and document-based information across 10 sectors.
Visual materials
Tasks involve blueprints, scans and charts, beyond text-only evaluation.
Expert-built tasks
Models must interpret information and complete demanding agent tasks.
Top MKA scores
The three leading results fall within the benchmark’s stated margin of error.
Access and pricing
Availability depends on the Google product and user account.
| Route | Who can access | API rates |
|---|---|---|
| Gemini API, AI Studio, Gemini for business | Paying users | $2 input / $10 output per million tokens through November 30 |
| Gemini app, AI Mode in Search | Google AI Ultra subscribers | From December 1: $4 input / $20 output per million tokens |
What buyers should weigh
A benchmark result is only one part of the deployment decision.
- Google reports a much lower cost per MKA task than Fable 5.2.
- Carbon is designed for complex, multi-step work and extended verification.
- MKA tasks are private, limiting independent reproduction.
- Real costs depend on workload, settings, retries and human review.
- Latency and reliability may affect day-to-day value.
The Price of Visual-Agent Work
The MKA result puts a price-performance question in front of businesses considering agents that must work with documents, diagrams and other visual information. If Google DeepMind’s reported costs hold up in comparable use, Carbon could be much cheaper to run on these tasks than Fable 5.2, while scoring close to it on this benchmark. That matters for organizations weighing model quality against the cost of processing many tasks.
The comparison is not a general guarantee of lower bills. The reported figures concern cost per MKA task, not every workload or a fixed amount of work across all applications. Carbon also uses more thinking tokens than its sibling model, Gemini 4 Argon, and Google lists higher token use as a limitation, particularly at higher thinking settings. Real costs will depend on task design, input size, model settings and how often a system needs retries or human review.
Performance rankings also depend on what is tested. Artificial Analysis ranks Carbon behind Fable 5.2 on its current index, while Google’s own selected scorecard shows Carbon leading most of its listed tests. For buyers, the useful takeaway is not that one model is best at everything, but that benchmarks, access conditions, speed and workload-specific costs all affect the decision.
Products Worth Considering
Sony WH-1000XM5 Premium Noise Cancelling Wireless Headphones – Midnight Blue
Jackery Explorer 300 Portable Power Station with 40W Solar Panel
EcoFlow Delta 3 Max Portable Power Station with 220W Solar Panel, 2048Wh
As an affiliate, we earn on qualifying purchases.
From Limited Access to Paid Release
Carbon was initially provided to a small group of cyber defenders through Google’s Fairwind Program. Artificial Analysis lists October 28 as its release date. The November 10 announcement expanded access to paid users, while Google’s security features for the model remain restricted to verified defenders, according to the source material.
Google describes Carbon as a higher-compute model built on Gemini 4 Argon, intended for demanding work such as complex multi-step coding, debugging and open-ended scientific research. Its model card says Carbon uses more thinking tokens to explore hypotheses and run extended verification. Google also says Argon remains the better choice for most software engineering and knowledge work, positioning Carbon as a specialist rather than a universal replacement.
The model accepts text, images, audio and video, and produces text. Its context window is up to 1 million tokens, with input and output counted together. Google lists low, medium and high thinking settings, with medium as the default. The model card gives a March 2026 knowledge cutoff and warns that knowledge may be less up to date in some areas.
““Argon remains the ideal choice for the majority of software engineering and knowledge work.””
— Google, in the Gemini 4 Carbon model card
Limits of the Published Comparisons
MKA was developed jointly by Google DeepMind and Artificial Analysis, and the tasks are not public. That means outside researchers cannot independently inspect the full task set or reproduce the reported results from the materials described here. The benchmark’s reported margin of error also means the small score differences among the top three should be read cautiously.
The cost figures are reported by Google DeepMind, the developer of one of the models being compared. The available material does not explain all assumptions behind the per-task calculation, such as how task complexity, retries or output length affect the bill. It is also unclear how closely MKA performance maps to results in particular workplace deployments.
Google’s model card acknowledges hallucinations, occasional slowness or timeouts, and higher token use. Artificial Analysis reports that Carbon took 25.6 minutes per task on its Intelligence Index evaluation, against a 12.6-minute median for comparable models. The published evidence therefore leaves open how its cost advantage on MKA balances against speed and reliability in day-to-day use.
Pricing Changes and Independent Tests
The next confirmed pricing milestone is December 1, 2026, when Google’s listed API rates are scheduled to double from the introductory $2 per million input tokens and $10 per million output tokens. Developers and businesses can now evaluate Carbon through the paid products listed by Google, subject to each product’s access terms.
Further comparison will depend on additional independent evaluations and on how the model performs in real deployments. Buyers will need to test their own tasks, measure total costs and account for latency, errors and any human review. The MKA scores provide an initial comparison, but the private task set and Google’s role in building the benchmark mean that the result is not, by itself, a complete measure of Carbon’s value.
Key Questions
Who can use Gemini 4 Carbon now?
Google says Carbon is available through its 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.
How did Carbon perform on MKA?
Carbon scored 54.3%, behind Claude Fable 5.2 at 56.6% and ahead of Claude Opus 5.5 at 53.0%. The three scores are reported to be within the benchmark’s margin of error.
Is Carbon one-sixth the price of Claude Fable 5.2?
Google DeepMind reported $5.63 per MKA task for Carbon and $35.28 for Fable 5.2, which is roughly a sixth as much on that benchmark. The figures are task-specific and do not establish that Carbon will cost one-sixth as much for other workloads.
When does Carbon’s introductory API pricing end?
Google’s listed introductory rates of $2 per million input tokens and $10 per million output tokens run through November 30, 2026. The scheduled rates from December 1 are $4 and $20 per million, respectively.
What are the main caveats about the MKA result?
Google DeepMind helped develop MKA, its task set is private, and Google reported the cost figures for its own model. The top three scores are also within the stated margin of error, so the benchmark does not show a decisive performance lead for Carbon or Fable.
Source: ThorstenMeyerAI.com
