🔍 Read the full analysis: Gemini 4 Carbon’s New MKA Standing And What It Costs on ThorstenMeyerAI.com
Get hardware and tech essentials 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 developers, businesses and Google AI Ultra subscribers on November 10, 2026, and published its model card. On the new MKA benchmark, Carbon scored 54.3%, behind Claude Fable 5.2 at 56.6%; Google DeepMind reported a lower cost per task for Carbon, though the test 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, 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 with 54.3%, behind Anthropic’s Claude Fable 5.2 at 56.6%, while Google DeepMind reported that Carbon cost far less per benchmark task.
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’s listed API price is $2 per million input tokens and $10 per million output tokens through November 30. From December 1, the rates rise to $4 per million input tokens and $20 per million output tokens.
On the new MKA benchmark, which covers 200 expert-built tasks in 10 sectors, Claude Fable 5.2 scored 56.6%, Carbon 54.3% and Claude Opus 5.5 53.0%. The source report says the top three results are within the benchmark’s margin of error. Google DeepMind reported an average cost of $5.63 per MKA task for Carbon, compared with $35.28 for Fable 5.2.
Google’s model card reports Carbon leading on 15 of 17 benchmarks it selected, but the comparison predates Fable 5.2 and is not the same as an independent ranking. Artificial Analysis’ Intelligence Index v5.0.1 gives Carbon, at its high setting, a score of 60. Fable 5.2 scored 62 at high and 65 at its maximum setting. Artificial Analysis also reported Carbon at $7.76 per task on its index and said it took 25.6 minutes per task, against a 12.6-minute median for comparable models.
Google describes Carbon as a model for demanding work, including multi-step coding and scientific research. Its model card says it uses more “thinking tokens” than Gemini 4 Argon to test possibilities and verify answers. Google says Argon remains the better fit for most software engineering and knowledge work. Carbon accepts text, images, audio and video, returns text, and has a context window of up to one million tokens, counting input and output together.
Gemini 4 Carbon performance and price
Google’s model placed second on a new document focused benchmark, with a lower reported task cost. The comparison has limits: Google helped develop the benchmark, and its tasks are private.
What the benchmark shows
MKA tests AI agents on expert built tasks involving documents such as blueprints, scans and charts.
Scored 56.6%, the highest result listed.
Scored 54.3%, with Google reporting lower cost per task.
Scored 53.0%. The top three results are within the reported margin of error.
API pricing changes
Google listed these rates for Gemini 4 Carbon.
| Period | Input per million tokens | Output per million tokens |
|---|---|---|
| Through November 30 | $2 | $10 |
| From December 1 | $4 | $20 |
Access depends on the product
- Gemini API
- Google AI Studio
- Gemini for business
- Gemini app access
- AI Mode in Search access
Fit and trade offs
Carbon is aimed at demanding work, while Google says Argon suits most software engineering and knowledge work.
Broad inputs
Accepts text, images, audio and video, with a context window of up to one million tokens.
Slower on another test
Artificial Analysis reported 25.6 minutes per Intelligence Index task, against a 12.6 minute median for comparable models.
Consider usage and reliability
Google documents higher token use at elevated thinking settings, occasional slowdowns or timeouts, and hallucinations.
Carbon’s Cost and Speed Trade-Off
The results give companies weighing high-end AI tools a comparison that includes both task performance and reported expense. Carbon’s MKA score trails Fable 5.2 by 2.3 percentage points, but Google DeepMind’s reported cost per task is substantially lower. Those figures may matter to organizations that process many complex visual or structured documents, although they come from a benchmark Google helped build.
Cost is not the only practical factor. Artificial Analysis characterized Carbon as slow per task and verbose, reporting 25.6 minutes for each Intelligence Index task. That is a different test from MKA, so its timing should not be treated as an MKA result. More broadly, Carbon’s own documentation lists higher token use at elevated thinking settings, occasional slowdowns or timeouts, and hallucinations. Buyers will need to weigh those limitations against the kinds of work they want the model to perform.
The timing also shows how quickly public model comparisons can change. Carbon had led an earlier version of the Artificial Analysis index, according to the source material, but Fable 5.2’s release on November 9 placed it ahead in the newer version. A benchmark position is a snapshot of particular models, settings and tests, not a guarantee that one model will perform best for every organization.
Products Worth Considering
Logitech MX Keys S Combo Wireless Keyboard and Mouse, Graphite
Apple MacBook Pro 16-inch (2021), M1 Pro, 10-core CPU, 16-core GPU, 16GB RAM, 512GB SSD, Silver
Trezor Safe 7 Crypto Hardware Wallet with Bluetooth and Color Touchscreen
As an affiliate, we earn on qualifying purchases.
From Defender 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 its broader paid release. Artificial Analysis lists October 28 as its release date. The source material says Carbon’s stronger security features remain restricted to verified defenders; general availability does not mean that all of those capabilities are open to every subscriber.
Google’s model card compares Carbon with Gemini 4 Argon, GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1. The card says its results are from October 2026, and it does not include Fable 5.2, which was released November 9. Some reported margins are narrow: Carbon scored 9.6% on SWE-sweep against 9.5% for GPT-6 Astra, and 70.4% on the Vals Index against 70.2% for Claude Opus 5.5. Other gaps were wider, including Carbon’s 96.2% on GraphWalks against 71.8% for the best rival listed there.
The model card also records areas where Carbon did not lead. GPT-6 Astra scored 65.5% on FrontierSWE 2, compared with Carbon’s 61.2%. Claude Opus 5.5 scored 87.2% on OSWorld-2.1, a desktop-operation test, while Carbon scored 81.5%. These results help explain why Google’s 15-of-17 figure should be read as a result on its chosen test set, rather than as a universal ranking.
““tasks that test the limits of a model’s intelligence””
— Google, in the Gemini 4 Carbon model card
Limits of the MKA Comparison
MKA was launched by Google DeepMind and Artificial Analysis, so Google helped develop the benchmark on which Carbon placed second. The task set is private, according to the source material, meaning outside researchers cannot inspect the questions or independently reproduce the published results. The reported ranking and prices are therefore useful data points, but their full methodology and generalizability cannot be checked from the information provided.
The source says the top three scores fall within the margin of error, but does not give the numerical margin or additional statistical details here. It is also not clear how closely MKA’s task mix represents the work of any particular company. The per-task cost figures are reported by Google DeepMind; they should not be taken as a guaranteed bill for a customer, since actual costs can vary with task, usage and model settings.
Some pricing and access details may change as Google’s availability terms evolve. The source gives the scheduled December 1 API price increase, but does not provide a later confirmation of whether the change has taken effect. It also does not specify when Carbon’s security capabilities might become available beyond verified defenders.
December Pricing and Further Testing
The next stated pricing milestone is December 1, 2026, when Google’s listed API rates are scheduled to move from $2 to $4 per million input tokens and from $10 to $20 per million output tokens. Developers and businesses considering Carbon will need to check Google’s current pricing and access terms before relying on those figures.
Further independent testing would help clarify how Carbon performs on MKA-like document tasks, particularly if researchers can evaluate a public or reproducible set of examples. Until then, the available picture combines Google’s model-card results, Artificial Analysis’ index and a new benchmark whose tasks remain private. The model’s performance, speed and cost in a particular customer’s workflow remain open questions.
Key Questions
Who can use Gemini 4 Carbon?
Carbon is available through the paid Gemini API, Google AI Studio and Gemini for business. In the Gemini app and AI Mode in Search, the source says access is limited to Google AI Ultra subscribers.
What did Carbon score on the MKA benchmark?
Carbon scored 54.3% on MKA, behind Claude Fable 5.2 at 56.6% and ahead of Claude Opus 5.5 at 53.0%. The source says the top three scores are within the margin of error.
How much does Gemini 4 Carbon cost?
Through November 30, Google’s listed API price is $2 per million input tokens and $10 per million output tokens. From December 1, the listed rates are $4 and $20, respectively. Actual use costs depend on token consumption and settings.
Is Carbon the top-ranked model?
Not on the cited current rankings. Carbon scored 60 on Artificial Analysis’ Intelligence Index v5.0.1, while Claude Fable 5.2 scored 62 at its high setting and 65 at its maximum setting. Google’s separate model card shows Carbon leading on 15 of 17 tests it selected, but that comparison does not include Fable 5.2.
Can the MKA results be independently checked?
Not fully from the information provided. Google DeepMind helped develop MKA, and its tasks are private, so outsiders cannot inspect or reproduce the full test set. The published scores and costs should be read with that limitation in mind.
Source: ThorstenMeyerAI.com
