🔍 Read the full analysis: Gemini 4 Carbon’s MKA Ranking Comes With A Smaller Price Tag on ThorstenMeyerAI.com
Get business pricing on tech for your team
- 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 AI Ultra subscribers on November 10, the same day it published results from the new MKA benchmark. Carbon ranked second at 54.3%, behind Claude Fable 5.2 at 56.6%, while Google DeepMind reported a much lower MKA cost per task: $5.63 versus $35.28. The benchmark was co-developed by Google DeepMind, its tasks are private, and the top scores fall within the stated margin of error.
Google opened Gemini 4 Carbon to paying developers, businesses and Google AI Ultra subscribers on November 10, the day Google DeepMind and testing company Artificial Analysis published results from their new MKA benchmark for agents working with visual documents. Carbon scored 54.3%, behind Anthropic’s Claude Fable 5.2 at 56.6%, but Google DeepMind’s report puts Carbon’s cost at $5.63 per task compared with $35.28 for Fable 5.2.
MKA evaluates AI agents on 200 expert-built tasks across 10 sectors, using material such as blueprints, scans and charts rather than only plain text. The reported scores put Claude Fable 5.2 first at 56.6%, Gemini 4 Carbon second at 54.3%, and Claude Opus 5.5 third at 53.0%. The source account says the three results are within the benchmark’s margin of error, so the ranking does not establish a clear performance gap among the leaders.
Access depends on the 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, it is restricted to Google AI Ultra subscribers. Google set introductory API prices at $2 per million input tokens and $10 per million output tokens through November 30. From December 1, those rates are scheduled to rise to $4 per million input tokens and $20 per million output tokens.
Google describes Carbon as its top-tier model for difficult work, including multi-step coding and open-ended research. Its model card says it uses more “thinking tokens” than Gemini 4 Argon to examine possibilities and verify answers. Google also says Argon remains the better fit for most software engineering and knowledge work. Carbon accepts text, images, audio and video, produces text, and has a context window of up to 1 million tokens, counting input and output together.
Gemini 4 Carbon’s MKA Ranking Comes With a Smaller Price Tag
Google opened Carbon to paying users as a new benchmark ranked it second on visual document tasks. Its reported cost per task was much lower than the top-scoring model’s.
What MKA measures
The benchmark tests agents on visual documents such as scans, charts and blueprints.
Document-based tasks
MKA uses 200 expert-built tasks across 10 sectors.
A close top three
Fable 5.2 scored 56.6%, Carbon 54.3%, and Opus 5.5 53.0%. The reported results fall within the margin of error.
Compare cost with care
The reported figures describe MKA tasks, not every company’s production costs.
- Lower reported cost per MKA task could matter for repeated document analysis.
- Carbon is available through paid developer and business offerings.
- Task length, token use and requested reasoning can affect spend.
- Google warns of higher token use and occasional slowness or timeouts.
Access and API pricing
Availability and rates depend on the product and date.
The Cost of Document-Based AI Tasks
The MKA result gives businesses a comparison that includes both task performance and reported cost. If the estimates hold for relevant workloads, Carbon’s lower per-task bill could matter to organizations that need agents to interpret diagrams, scans or charts repeatedly. That is a different purchasing question from simply choosing the model with the highest score.
The cost figures are tied to MKA’s tasks and the report’s evaluation, not a guarantee of what every customer will pay in production. API rates, task length, usage patterns and the amount of reasoning requested can affect spending. Google’s model card also warns of higher token use at higher thinking levels, as well as occasional slowness or timeouts. A low benchmark cost should therefore be weighed against task completion, latency and reliability in a company’s own tests.
Carbon’s second-place result also illustrates how quickly model comparisons can change. The source material says it had led a prominent independent leaderboard two weeks earlier, before broader access and the appearance of newer competitors. The new MKA score adds evidence about a particular kind of agent work; it does not settle which model is best across coding, desktop use, research or general assistance.
Products Worth Considering
Bose QuietComfort Wireless Noise-Cancelling Headphones, Sandstone
Samsung Galaxy Tab S7 11-inch Android Tablet 128GB Wi‑Fi with S Pen Mystic Black
Amazon Kindle Paperwhite (2024, 16 GB) – 7-inch Glare-Free Display, Black
As an affiliate, we earn on qualifying purchases.
From Restricted Preview to Paid Access
Carbon was initially made available to a small group of cyber defenders through Google’s Fairwind Program. Artificial Analysis lists October 28 as its release date. Google’s stronger security features for the model remain limited to verified defenders, according to the supplied source material; general paid access does not mean those features are open to all users.
Google’s November model card reports Carbon as the leading model on 15 of 17 benchmarks that Google selected for its comparison. That table includes Gemini 4 Argon, OpenAI’s GPT-6 Astra, and Anthropic’s Claude Opus 5.5 and Fable 5.1, but not Fable 5.2, which the source says was released November 9. The results are dated October 2026. Carbon’s leads vary: it scored 96.2% on GraphWalks against GPT-6 Astra’s 71.8%, while GPT-6 Astra led on FrontierSWE 2 and Claude Opus 5.5 led on OSWorld-2.1.
A separate Artificial Analysis index offers another comparison. In its v5.0.1 results, Carbon at the high thinking setting scored 60; Claude Fable 5.2 scored 62 at high and 65 at its maximum setting. The supplied account also reports that Carbon is slower per task on that index: 25.6 minutes compared with a 12.6-minute median for comparable models. Those figures concern a different evaluation from MKA and should not be treated as interchangeable.
““tasks that test the limits of a model’s intelligence””
Benchmark Limits and Pricing Questions
MKA was co-built by Google DeepMind, and the benchmark tasks are private, according to the supplied report. Those facts limit outside researchers’ ability to inspect the task set or independently reproduce the results. Google’s involvement does not by itself invalidate the scores, but it is relevant when interpreting a benchmark on which Google’s model performs strongly.
The reported difference between Carbon and Fable 5.2 is also small in score terms, and the source says the top three models are within the margin of error. The material does not provide the size of that margin or enough detail to judge how results vary across sectors and task types. It is therefore unclear whether the order would hold in a larger or independently replicated evaluation.
The $5.63 and $35.28 figures are reported costs per MKA task, not universal prices for comparable customer work. The supplied material does not specify enough about how those costs were calculated to map them directly to a company’s expected bill. Carbon’s long completion times on another evaluation and Google’s warnings about token use are further factors buyers may need to test.
Access and Rate Changes Ahead
Developers and businesses can evaluate Carbon through the paid Gemini API, Google AI Studio and Gemini for business, while Gemini app and Search AI Mode access remains tied to Google AI Ultra. The next stated pricing milestone is December 1, when the introductory API rates are scheduled to double from $2/$10 to $4/$20 per million input/output tokens.
For organizations considering the model, the immediate practical next step is to test their own workloads rather than rely on a single benchmark score. Relevant checks include answer quality on their documents, completion time, token consumption, timeout frequency and total cost per successfully completed task. Further MKA detail or independent evaluation could help clarify how broadly the published ranking applies, but the supplied material does not announce a date for additional benchmark results.
Key Questions
What was the main Gemini 4 Carbon news?
Google made Gemini 4 Carbon available to paying users on November 10, 2026, and published results from the new MKA benchmark. Carbon ranked second, behind Claude Fable 5.2.
How did Carbon score on MKA?
Carbon scored 54.3%. Claude Fable 5.2 scored 56.6%, and Claude Opus 5.5 scored 53.0%. The source says the top three results are within the margin of error.
Why is Carbon described as cheaper?
Google DeepMind’s report puts Carbon’s MKA cost at $5.63 per task, compared with $35.28 for Fable 5.2. That is a benchmark-specific estimate, not a price guarantee for every customer workload.
Who can use Gemini 4 Carbon?
It is available through the paid Gemini API, Google AI Studio and Gemini for business. Access in the Gemini app and AI Mode in Search is limited to Google AI Ultra subscribers.
What happens to Carbon’s API price in December?
Introductory pricing runs through November 30, 2026. From December 1, Google’s listed rates are scheduled to be $4 per million input tokens and $20 per million output tokens, up from $2 and $10.
Source: ThorstenMeyerAI.com
