🔍 Read the full analysis: Gemini 4 Carbon At A Sixth Of The Leader’s Cost: What To Know 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, 2026. Carbon ranked second on the new MKA benchmark, behind Claude Fable 5.2, while Google DeepMind reported a much lower per-task cost; the benchmark was co-developed by Google and its tasks are private.
Google opened Gemini 4 Carbon to paying developers, businesses and Google AI Ultra subscribers on November 10, 2026, the same day Google DeepMind and testing firm Artificial Analysis introduced a benchmark for AI agents working with visual documents. Carbon placed second, at 54.3%, behind Anthropic’s Claude Fable 5.2 at 56.6%; Google DeepMind reported Carbon’s cost per benchmark task at $5.63, compared with $35.28 for the leader.
The MKA benchmark comprises 200 expert-built tasks across 10 sectors. It tests agents on materials such as blueprints, scans and charts, rather than relying only on plain text. Claude Opus 5.5 scored 53.0%, putting the three leading models close together. The source report says their scores fall within the margin of error, so the ordering should not be read as a definitive measure of overall capability.
Google describes Carbon as its top-tier Gemini 4 model, built for demanding work such as multi-step coding and open-ended scientific research. Its model card says it can take text, images, audio and video as input and has a context window of up to one million tokens. Google also says Carbon uses more “thinking tokens” than Gemini 4 Argon, which can mean higher usage and slower responses.
Access differs by product. Developers can use Carbon through the Gemini API and Google AI Studio, while businesses can access it through Gemini for business. In the Gemini app and AI Mode in Search, it 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; from December 1, those rates rise to $4 and $20.
Gemini 4 Carbon at a lower cost What businesses should know
Carbon ranked second on a new visual-document benchmark, with a reported per-task cost far below the leader. The comparison comes with limits on what the score and cost can tell buyers.
What the benchmark found
The top three scores are close and fall within the reported 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 stated |
Why the comparison matters
MKA tests agents on visual documents such as blueprints, scans and charts.
- Lower reported cost could matter for organizations running many agent tasks.
- Carbon is designed for demanding work, including multi-step coding and scientific research.
- The cost figure applies to MKA tasks, not every customer workload.
- Artificial Analysis describes Carbon as slow per task and verbose.
- Google lists hallucinations, occasional timeouts and higher token use as limitations.
Access and API pricing
Availability depends on the Google product and subscription.
Paid access through the Gemini API, Google AI Studio and Gemini for business.
Access is limited to Google AI Ultra subscribers.
Through November 30: $2 per million input tokens and $10 per million output tokens. From December 1: $4 and $20.
Read the rankings with care
The score is one snapshot, shaped by its tasks, settings and test selection.
Cost and Speed Shape Carbon’s Appeal
The result presents Carbon as a possible lower-cost option for companies that need an agent to interpret complex visual documents, but the benchmark does not show that it is the strongest model on every task. Its 54.3% MKA score trails Fable 5.2’s 56.6%, and the reported cost comparison applies to these benchmark tasks, not every customer workload.
Google DeepMind’s report puts Carbon at $5.63 per MKA task and Fable 5.2 at $35.28—roughly one-sixth as much. If the costs are comparable in real deployments, that gap could matter to organizations running many agent tasks. However, the source material does not provide enough detail to treat the benchmark’s cost figure as a forecast of a company’s actual bill.
Speed is another trade-off. Artificial Analysis describes Carbon as slow per task and verbose, reporting an average of 25.6 minutes per task on its Intelligence Index, against a 12.6-minute median for comparable models. That is a separate test from MKA and points to a practical question for users: lower task cost may not mean faster completion.
Products Worth Considering
Jackery Explorer 300 Portable Power Station with 40W Solar Panel
Logitech MX Mechanical Wireless Illuminated Keyboard, Tactile Quiet, Graphite
Samsung Galaxy Tab A11+ 8GB RAM, 256GB Storage, Gray
As an affiliate, we earn on qualifying purchases.
How the Rankings Changed
Carbon was initially made available to a small group of cyber defenders through Google’s Fairwind Program. Artificial Analysis lists its release date as October 28, about two weeks before the broader paid rollout. Google says enhanced security features remain limited to verified defenders.
Google’s model card gives Carbon the best result in 15 of 17 benchmarks that Google selected for comparison. Those results are dated October 2026 and do not include Claude Fable 5.2, which the source says was released November 9. Carbon’s leads also vary: it narrowly edges rivals on some tests, while scoring 96.2% on GraphWalks, compared with 71.8% for the best rival listed there. GPT-6 Astra leads Carbon on FrontierSWE 2, and Claude Opus 5.5 leads on OSWorld-2.1.
The independent Artificial Analysis Intelligence Index offers a different comparison. In version 5.0.1, Carbon at its high thinking setting scores 60 points; Fable 5.2 scores 65 at its maximum setting and 62 at high. The index combines nine tests, and its rankings depend on the evaluated settings and test selection. The differing snapshots help explain why a model can lead a company’s chosen scorecard but rank second on a newer outside comparison.
““tasks that test the limits of a model’s intelligence””
— Google, in the Gemini 4 Carbon model card
Benchmark Limits and Open Questions
MKA was co-developed by Google DeepMind, whose own model ranked second, and its tasks remain private. That means outside researchers cannot inspect the task set or independently reproduce the scores from the materials described in the source. The reported margin of error also leaves the precise ordering of the top three uncertain.
It is not clear from the source how representative MKA’s 200 tasks are of routine commercial work, or how its per-task cost calculation maps to the API bills customers will receive. The reported score and cost do not establish that Carbon will be cheaper or more effective for every workflow.
Google lists hallucinations, occasional slowness or timeouts, and higher token use—especially at higher thinking levels—as limitations. The available information does not give a complete account of how often those issues occur under different workloads. The model card’s benchmark results also use a selected set of tests, and the cited snapshot predates Fable 5.2’s release.
Access, Pricing and Independent Testing
Developers and businesses can now try Carbon through Google’s paid offerings, while app and Search access remains tied to Google AI Ultra. Users considering API adoption will face the published introductory rates through November 30, 2026, followed by the higher listed rates from December 1.
The next useful evidence will be additional testing of Carbon on publicly inspectable tasks, along with independent evaluations that include Fable 5.2 and clarify how model settings affect results. For businesses, real-world measures of accuracy, completion time, token use and total cost will matter alongside headline benchmark rankings. The source material does not identify a date for a later benchmark update or further access expansion.
Key Questions
Who can use Gemini 4 Carbon now?
Paid developers and businesses can access it through Google’s developer and business offerings. In the Gemini app and AI Mode in Search, access is limited to Google AI Ultra subscribers.
Did Carbon win the new MKA benchmark?
No. Claude Fable 5.2 scored 56.6%, ahead of Carbon at 54.3%. Claude Opus 5.5 scored 53.0%. The source report says the top three are within the margin of error.
Why is Carbon described as costing about one-sixth of the leader?
Google DeepMind reported $5.63 per MKA task for Carbon, compared with $35.28 for Fable 5.2. This is a benchmark-specific cost comparison, not a guaranteed price for other customer workloads.
How much does Gemini 4 Carbon cost through the API?
The listed price through November 30, 2026, is $2 per million input tokens and $10 per million output tokens. From December 1, the listed rates rise to $4 and $20, respectively.
What are the main drawbacks Google identifies?
Google lists hallucinations, occasional slowness or timeouts, and higher token use, particularly at higher thinking levels. Artificial Analysis also reported a long average task time on its separate Intelligence Index evaluation.
Source: ThorstenMeyerAI.com
