🔍 Read the full analysis: Gemini 4 Carbon Is Here With A Lower-Cost, Second-Place MKA Result on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Google opened Gemini 4 Carbon to paying developers, businesses and Google AI Ultra subscribers on November 10, and published its model card. Carbon scored 54.3% on the new MKA benchmark, behind Claude Fable 5.2 at 56.6%, while Google DeepMind reported a much lower cost per MKA task for Carbon; the benchmark’s tasks are private and Google helped build the test.
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 documents such as blueprints, scans and charts. Carbon placed second, at 54.3%, behind Anthropic’s Claude Fable 5.2 at 56.6%; Google DeepMind reported that Carbon’s cost per task on the test was substantially lower.
The new test, called MKA, contains 200 expert-built tasks spanning 10 sectors. Claude Fable 5.2 scored 56.6%, Gemini 4 Carbon 54.3%, and Claude Opus 5.5 53.0%. The report says the top three scores are within the margin of error, so the ranking does not establish a clear performance gap among them.
Google DeepMind’s report puts Carbon’s average MKA cost at $5.63 per task, compared with $35.28 for Fable 5.2. That price comparison is specific to the benchmark’s reported task costs; it does not establish that Carbon is cheaper for every workload. The tasks are private, limiting independent scrutiny of how the benchmark was constructed and how well the results generalize.
Carbon is available through the paid Gemini API and Google AI Studio, as well as Gemini for business. In the Gemini app and AI Mode in Search, access is limited to Google AI Ultra subscribers. Google lists an introductory price through November 30 of $2 per million input tokens and $10 per million output tokens. From December 1, the listed rates rise to $4 and $20 per million tokens, respectively.
Gemini 4 Carbon Is Here Lower Cost, Second Place
Google opened Carbon to paying developers and businesses on November 10. It ranked second on the new MKA document-agent benchmark, with a reported per-task cost below Claude Fable 5.2.
What MKA measures
The benchmark tests agents on visual and structured documents, including blueprints, scans and charts.
200 expert-built tasks
Tasks span 10 sectors and focus on document-based work.
Close results
Claude Opus 5.5 scored 53.0%. The top three are within the reported margin of error.
Who can access Carbon
Availability depends on the Google product and subscription.
- Developers through the Gemini API and Google AI Studio
- Businesses through Gemini for business
- Gemini app access requires Google AI Ultra
- AI Mode in Search access requires Google AI Ultra
Token pricing changes
Google lists introductory rates through November 30, then higher rates from December 1.
| Period | Input per million tokens | Output per million tokens |
|---|---|---|
| Through November 30 | $2 | $10 |
| From December 1 | $4 | $20 |
What the result does not settle
The reported benchmark cost and score are specific to MKA; they do not predict every organization’s workload.
- MKA tasks are private, limiting independent scrutiny and reproduction.
- The reported cost does not establish lower costs for every use case.
- Latency, reliability, token use and human review also affect value.
Cost and Performance for Agent Tasks
The MKA result gives businesses a new, task-specific comparison for models expected to interpret visual and structured documents, rather than just answer questions about plain text. That can matter in work involving technical diagrams, scanned forms and charts, where an agent must extract information and act on it. Carbon’s second-place score suggests it is competitive on this test, but the narrow gaps and stated margin of error counsel against treating the ranking as a definitive ordering.
The reported cost difference may be relevant to organizations evaluating repeated or large-scale agent workflows. On MKA, Google DeepMind says Carbon cost $5.63 per task versus $35.28 for Fable 5.2. But price alone does not show which model offers better value for a particular job: performance, latency, reliability and the amount of human review required also affect total cost. Artificial Analysis separately describes Carbon as slow and verbose on its Intelligence Index, reporting 25.6 minutes per index task against a 12.6-minute median for comparable models.
Availability also has limits. Carbon is not being offered to all users of Google’s consumer services: access in the Gemini app and Search’s AI Mode requires Google AI Ultra. Its broader release is most directly relevant to paying developers and organizations able to test it through Google’s developer and business products.
Products Worth Considering
Amazon Kindle Paperwhite (2024, 16 GB) – 7-inch Glare-Free Display, Black
Bose TV Speaker (2nd Gen) Compact Soundbar
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.
A Specialist Model With Longer Reasoning
Google describes Carbon as its top-tier model for demanding work, including multi-step coding and debugging and open-ended scientific research. It is built on Gemini 4 Argon and, according to Google’s model card, uses more thinking tokens to explore hypotheses and run extended verification. Google says Argon remains the better choice for most software engineering and knowledge-work tasks, positioning Carbon as a model for harder cases rather than a universal replacement.
The model card lists support for text, images, audio and video inputs, with text output, and a context window of up to 1 million tokens for input and output combined. Google gives its knowledge cutoff as March 2026, while cautioning that knowledge may be less current in some areas. The model has low, medium and high thinking settings, with medium as the default. Google also lists hallucinations, occasional slowness or timeouts, and higher token use—particularly at higher thinking settings—as limitations.
Carbon had earlier been made available to a small group of cyber defenders through Google’s Fairwind Program, about two weeks before the broader paid release, according to the supplied source. Its stronger security features remain limited to verified defenders. In Google’s model-card comparison, Carbon led 15 of 17 selected benchmarks, but that table used Google’s chosen tests and rivals and did not include Fable 5.2, released November 9. An independent Artificial Analysis index gives a different snapshot: Carbon scored 60 in version 5.0.1, while Fable 5.2 scored 62 at the high setting and 65 at its maximum.
““tasks that test the limits of a model’s intelligence””
— Google, in the Gemini 4 Carbon model card
Limits of the MKA Comparison
The MKA tasks are not publicly available, and Google DeepMind helped create the benchmark. That does not by itself invalidate the scores, but it means outside researchers cannot directly inspect the tasks or independently reproduce the comparison from the information provided. The source material does not provide enough detail to assess the full test methodology or the precise uncertainty range behind the reported margin of error.
It is also unclear how closely MKA performance predicts results on real business workloads. A score across 200 tasks cannot establish how a model will perform on every sector’s documents, or account for differences in deployment, prompt design and human oversight. The reported cost per task likewise should not be treated as a general operating-cost estimate.
Google’s benchmark table and the independent index measure different things and use different model selections and settings. The model card’s results are dated as of October 2026, while the current Artificial Analysis index includes Fable 5.2. The supplied figures do not provide a single, like-for-like comparison across all models, tasks, settings, speed and cost.
Pricing Changes in December
Developers and organizations can test Carbon through Google’s paid API, AI Studio and business offerings, subject to their account access. Google has listed introductory token rates through November 30, 2026; the listed rates change on December 1. Buyers comparing models will need to assess their own task quality, latency and token use rather than relying on MKA rank or cost alone.
Further independent testing could clarify how Carbon performs on publicly inspectable tasks and on deployed workflows. The provided material does not identify a scheduled release of MKA’s private tasks or a date for another benchmark update. For now, the confirmed picture is a paid release, a second-place MKA score within the stated margin of error, and a reported per-task cost advantage on that benchmark—not a settled verdict on which model is best for general use.
Key Questions
Who can use Gemini 4 Carbon now?
Paying developers and businesses can access it through the Gemini API, Google AI Studio and Gemini for business. Use in the Gemini app and AI Mode in Search is restricted to Google AI Ultra subscribers.
Did Carbon win the MKA benchmark?
No. Carbon scored 54.3%, behind Claude Fable 5.2 at 56.6%. The report says the top three results fall within the margin of error, so the close ranking should be read cautiously.
How much does Carbon cost per MKA task?
Google DeepMind reported an average of $5.63 per MKA task for Carbon, compared with $35.28 for Fable 5.2. Those are benchmark-specific figures, not a general price guarantee for other workloads.
What are Carbon’s listed token prices?
Google lists introductory rates through November 30, 2026, of $2 per million input tokens and $10 per million output tokens. From December 1, the listed rates are $4 and $20 per million tokens.
What is still uncertain about the benchmark?
The MKA tasks are private, and Google DeepMind helped build the test. That limits outside inspection and reproduction; the available information also does not establish how closely scores predict performance on every real-world workflow.
Source: ThorstenMeyerAI.com
