🔍 Read the full analysis: The MKA Test Puts Gemini 4 Carbon In Second Place 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 made Gemini 4 Carbon available to paying users on November 10, the same day Google DeepMind and Artificial Analysis introduced the MKA benchmark. Carbon scored 54.3%, placing second behind Claude Fable 5.2’s 56.6%; Google DeepMind reported a much lower cost per MKA task for Carbon, while the benchmark’s private tasks and Google’s role in its development limit independent scrutiny.
Google opened Gemini 4 Carbon to paying developers, businesses and Google AI Ultra subscribers on November 10, the same day Google DeepMind and testing firm Artificial Analysis released the new MKA benchmark for AI agents handling materials such as blueprints, scans and charts. Carbon scored 54.3%, 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 comprises 200 expert-built tasks across 10 sectors, according to the source report. Claude Opus 5.5 scored 53.0%, putting the top three results close together. The report says the differences among those scores are within the margin of error, so the ranking should not be read as proof of a clear performance gap. The tasks are private, meaning outside researchers cannot inspect or independently reproduce them.
Carbon is available through the paid Gemini API, Google AI Studio and Gemini for business. In Google’s Gemini app and AI Mode in Search, access is limited to Google AI Ultra subscribers. The listed API price is $2 per million input tokens and $10 per million output tokens through November 30, rising to $4 and $20, respectively, on December 1, according to the source. Those rates are separate from the MKA cost-per-task estimates, which depend on benchmark usage.
Google describes Carbon as a high-end model for demanding tasks, including complex coding and scientific research. Its model card says it uses more thinking tokens than Gemini 4 Argon to work through problems and verification. Google also lists hallucinations, occasional slowness or timeouts, and higher token use among its limitations. The model accepts text, images, audio and video, returns text, and has a combined input-and-output context window of up to one million tokens.
Gemini 4 Carbon Places Second on the MKA Test
Carbon scored close to the top, with a lower reported task cost. Private test tasks and close scores limit what buyers can conclude.
What the ranking shows
The top three results are reported within the margin of error.
A close second
Carbon scored below Fable 5.2 and above Claude Opus 5.5 at 53.0%. The small gaps do not establish a clear performance lead.
Lower benchmark estimate
Google DeepMind reports a lower cost per MKA task for Carbon. Actual customer costs depend on usage and workflow.
Access and pricing
Availability depends on how people use Carbon.
- Gemini API and Google AI Studio
- Gemini for business
- Gemini app and Search AI Mode for Ultra subscribers
- Through November 30: $2 input and $10 output per million tokens
- Scheduled from December 1: $4 input and $20 output per million tokens
- Token rates differ from benchmark task-cost estimates
Limits on independent checks
MKA covers agent tasks involving materials such as blueprints, scans, and charts.
Outside researchers cannot inspect or independently reproduce the test items.
That is relevant context when assessing results, but does not by itself show bias.
Benchmark scores do not establish performance or operating costs for a company’s own work.
What buyers should weigh
Speed, access, and task fit can matter alongside the headline score.
| Consideration | What the report says |
|---|---|
| Task speed | 25.6 minutes per task on Artificial Analysis’ index, versus a 12.6-minute median for comparable models |
| Reasoning use | Google says Carbon uses more thinking tokens than Gemini 4 Argon |
| Other limitations | Google lists hallucinations, occasional slowness or timeouts, and higher token use |
Performance Comes With a Cost Trade-Off
The MKA result gives companies a fresh comparison for AI agents expected to interpret visual or structured information, rather than answer questions from text alone. For buyers, the figures suggest a possible trade-off: Carbon scored slightly below Fable 5.2 in this test, while the report attributes a substantially lower per-task cost to Carbon. That may matter for organizations running many agent tasks, though benchmark costs do not by themselves predict a customer’s bill or performance on its own workflows.
The result also complicates the picture presented by Google’s own model card. That card reports Carbon leading on 15 of 17 benchmarks selected by Google, but it does not include Claude Fable 5.2, which the source says was released November 9. Meanwhile, Artificial Analysis’ separate Intelligence Index v5.0.1 places Carbon at 60, below Fable 5.2 at 65 at maximum settings and 62 at high settings. These are different tests with different methods; none alone establishes which model is best for every use.
Access and price are practical considerations for individual users and organizations. Carbon’s consumer-facing availability is restricted to Ultra subscribers, while API and business access is paid. Its standard API rates are scheduled to rise on December 1, and the source says Artificial Analysis found it notably slow per task on its own index: 25.6 minutes per task, compared with a 12.6-minute median for comparable models. Speed, access conditions and the type of work may matter as much as a headline ranking.
Products Worth Considering
Jackery Explorer 300 Portable Power Station with 40W Solar Panel
Apple AirPods Pro (2nd Generation) Wireless Earbuds with USB‑C Charging, Up to 2X More Active Noise Cancelling, Transparency Mode, Adaptive
Sonos Arc Ultra Soundbar with Dolby Atmos, Black
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 wider paid release, according to the source. Artificial Analysis lists October 28 as its release date. Its stronger security features remain limited to verified defenders, the source reports; general paid access should not be confused with access to those restricted capabilities.
Google’s model card compares Carbon with Gemini 4 Argon, OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5.5 and Fable 5.1. Google says its results are as of October 2026. In that selected set, Carbon leads 15 of 17 entries, but some margins are narrow: it scored 9.6% against GPT-6 Astra’s 9.5% on SWE-sweep and 70.4% against Claude Opus 5.5’s 70.2% on the Vals Index. It trails on FrontierSWE 2 and OSWorld-2.1.
The independent Artificial Analysis index offers another snapshot. In version 5.0.1, Carbon at high thinking scored 60; Fable 5.2 scored 62 at high and 65 at its maximum setting. The prior model-card comparison used version 5.0, where Carbon had led Claude Opus 5.5 by one point. Benchmark versions and model settings differ, so those scores are not interchangeable.
Private Tasks Limit Outside Checks
The MKA tasks are not publicly available, according to the source report. That prevents independent researchers from checking the test items or reproducing the reported results. Google DeepMind helped develop the benchmark, and Google’s model placed well on it; that is relevant context for readers, but it does not by itself show that the results are biased. Independent validation remains limited.
The published scores also do not establish how the models compare across a customer’s own documents, workflows, or operating costs. The top three MKA results are reported as being within the margin of error, and the source does not provide enough detail here to assess how individual sectors or task types affected each model’s score. The exact impact of longer reasoning, token use and latency on real-world bills is also unclear.
Pricing terms and access can change, and the scheduled December 1 rate increase has not yet taken effect as of November 11. The source does not establish whether Google will change those prices, expand app access, or make Carbon’s restricted security features available to a wider group.
Watch Pricing and Independent Replication
The next dated change in the source is Google’s scheduled API price increase on December 1, 2026, from $2 to $4 per million input tokens and from $10 to $20 per million output tokens. Customers considering Carbon will need to distinguish those token rates from estimated benchmark task costs and test the model against their own workloads.
Further comparisons may clarify how Carbon performs as additional independent evaluations and real-world usage data become available. A public release of MKA tasks or a separate replication would allow outsiders to examine the benchmark more closely, but the source reports no such release or timetable. For now, the published evidence supports a measured conclusion: Carbon is competitive on MKA, but the close scores, private tasks and differences among benchmarks leave its broader standing unsettled.
Key Questions
What was Gemini 4 Carbon’s score on the MKA test?
Carbon scored 54.3% on MKA, according to the source report. 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.
Why did Carbon place second?
Its reported score was below Claude Fable 5.2’s 56.6% and above Claude Opus 5.5’s 53.0%. Because the reported differences are within the margin of error, second place does not establish a decisive performance advantage for either of the top two models.
How much does Carbon cost to use?
The source reports API prices of $2 per million input tokens and $10 per million output tokens through November 30, 2026. From December 1, those rates are scheduled to become $4 and $20. Google DeepMind separately reported an MKA cost of $5.63 per task; that is a benchmark estimate, not a universal fee.
Can the MKA results be independently checked?
Not fully based on the information reported: the 200 MKA tasks are private, so outsiders cannot inspect or reproduce the test directly. Google DeepMind also co-developed the benchmark. Those facts limit scrutiny, but do not by themselves establish that the results are unreliable.
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, access is limited to Google AI Ultra subscribers, according to the source report.
Source: ThorstenMeyerAI.com
