🔍 Read the full analysis: Can Gemini 4 Carbon Deliver Strong MKA Results For Less? 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 opened Gemini 4 Carbon to paying developers, businesses and Google AI Ultra subscribers on Nov. 10, the day a new benchmark put it second to Claude Fable 5.2 on tasks involving visual documents. Google DeepMind reports Carbon costs $5.63 per MKA task, compared with $35.28 for Fable 5.2, but the benchmark was co-developed by Google and its private tasks cannot be independently checked.
Google opened Gemini 4 Carbon to paying developers, businesses and Google AI Ultra subscribers on Nov. 10, the same day Google DeepMind and testing firm Artificial Analysis introduced the MKA benchmark for AI agents working with blueprints, scans and charts. Carbon scored 54.3%, placing second behind Anthropic’s Claude Fable 5.2 at 56.6%, while Google DeepMind’s report put Carbon’s cost at $5.63 per task versus $35.28 for the leader.
The MKA results also placed Claude Opus 5.5 third at 53.0%. The benchmark comprises 200 expert-built tasks across 10 sectors. The reported scores are close, and the source account says the top three are within the margin of error; the results do not establish a clear performance separation. Google DeepMind’s cost figures make the comparison notable, but they are reported estimates tied to this benchmark, not a general measure of what every customer will pay.
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 set introductory API pricing at $2 per million input tokens and $10 per million output tokens through Nov. 30. From Dec. 1, those rates are scheduled to rise to $4 and $20, respectively.
Google’s model card gives Carbon the top score on 15 of 17 benchmarks selected by Google. Those results have limits as a comparison: the company chose the tests and competitors, the scores are dated October 2026, and Claude Fable 5.2, released Nov. 9, is not included. Carbon narrowly leads on some tests, including SWE-sweep, where it scores 9.6% against GPT-6 Astra’s 9.5%; it trails on FrontierSWE 2 and OSWorld-2.1.
Can Gemini 4 Carbon Deliver Strong MKA Results for Less?
Gemini 4 Carbon ranked close to the leader on visual-document tasks, with a much lower reported cost per task. The results are benchmark-specific and difficult to verify independently.
Close scores, lower reported cost
The top three results fall within the reported margin of error, so the ranking does not establish a clear performance gap.
What businesses should weigh
MKA tests AI agents on visual and technical materials such as blueprints, scans and charts.
Promising benchmark estimate
Carbon’s reported task cost is lower, but it is not a universal customer price or the same as API token pricing.
Test your own documents
A close MKA score does not show how reliably Carbon handles a company’s workflows.
Allow for longer tasks
Carbon is described as slow per task, and higher thinking levels can use more tokens.
Access and API pricing
Paying developers and businesses can access Carbon; Gemini app and Search AI Mode access is limited to Ultra subscribers.
Limits on the evidence
The results offer a useful signal, but not a complete verdict on model quality or customer costs.
- Google co-developed the benchmark.
- Its tasks are private, making independent replication difficult.
- The reported margin of error blurs the small score differences.
- Google selected the benchmarks and competitors.
- Scores are dated October 2026.
- Claude Fable 5.2, released Nov. 9, is not included.
Cost Versus MKA Performance
For companies considering AI agents that must interpret visual and technical material, MKA offers a comparison centered on tasks beyond plain-text prompts. Carbon’s second-place score suggests it performed near the leading result in this test, while the reported per-task cost could make it worth evaluating for workloads where price matters. The benchmark alone does not show how the models will perform on a particular company’s documents or processes.
The price comparison also has boundaries. MKA’s $5.63 figure is not the same as the API’s per-token list price, and task costs can depend on the task and model settings. Carbon can use more tokens when asked to think longer, and Google lists higher token use and occasional delays among its limitations. Buyers would need to test their own workloads and account for speed, accuracy and usage before estimating costs.
There is also a transparency issue: Google co-developed the benchmark, and the tasks remain private. That does not invalidate the published results, but it makes independent replication difficult. Readers should treat the ranking and cost figures as useful reported evidence, not a complete or independently verified verdict on model quality.
Products Worth Considering
Ledger Nano X Bluetooth Crypto Wallet
Apple MacBook Air 15-inch (2026) with M5 Chip, 16GB Memory, 1TB SSD, and Liquid Retina Display – Sky Blue
LG UltraGear 32GX870A-B 32-inch 4K OLED Gaming Monitor
As an affiliate, we earn on qualifying purchases.
Carbon’s Release and Model Profile
Carbon is Google’s high-end model built on Gemini 4 Argon. Google describes it as intended for demanding work, including complex multi-step coding and open-ended scientific research. Its model card says it uses more thinking tokens than Argon to explore hypotheses and carry out extended verification. Google also says Argon remains the better fit for most software engineering and knowledge-work tasks, positioning Carbon as a specialist rather than a replacement for routine use.
The model accepts text, images, audio and video, and produces text. Its context window supports up to 1 million tokens across input and output, and its listed knowledge cutoff is March 2026, though Google cautions that knowledge can be less current in some fields. Developers can select low, medium or high thinking levels; medium is the default. Google lists hallucinations, occasional slowness or timeouts, and increased token use at higher thinking levels as known limitations.
Carbon had previously been made available to a small group of cyber defenders through Google’s Fairwind Program. Artificial Analysis lists Oct. 28 as its release date. Its stronger security features remain restricted to verified defenders, according to the source material. The Nov. 10 opening therefore expanded access, but not every feature or access route is available to the general public.
““tasks that test the limits of a model’s intelligence””
— Google, in the Gemini 4 Carbon model card
Benchmark Limits and Open Questions
The MKA tasks are private, leaving outside researchers unable to reproduce the test directly. Google DeepMind co-built the benchmark, and the available account does not provide enough detail to independently assess how the task mix or scoring might affect model rankings. The reported margin of error also means the 2.3-point gap between Carbon and Fable 5.2 should not be read as a decisive advantage for either model.
It remains unclear how well MKA performance transfers to real customer work, including the accuracy and reliability of outputs on a company’s own blueprints, scans or charts. The supplied material also does not specify enough about the cost calculation to treat $5.63 and $35.28 as universal prices for comparable deployments. Separately, Google’s model-card results use a company-selected test set, and their October cutoff predates the release of Fable 5.2.
Carbon’s speed may also matter to users comparing costs. Artificial Analysis describes the model as slow per task and verbose, reporting 25.6 minutes per index task against a 12.6-minute median for comparable models. That measurement is specific to its index evaluation; it does not establish how long Carbon will take on every workload.
Access, Pricing and Further Testing
Developers and businesses can evaluate Carbon through Google’s paid developer and business offerings, while Gemini app and Search AI Mode access remains limited to Ultra subscribers. The next stated pricing date is Dec. 1, 2026, when API rates are scheduled to move from the introductory $2/$10 per-million-token prices to $4/$20. Users assessing the model should distinguish that token pricing from benchmark-specific cost-per-task estimates.
Further comparison will depend on additional independent testing, more detail on MKA’s evaluation method, and reports from customers using Carbon on their own tasks. Until those are available, the published numbers support a narrow conclusion: Carbon scored close to the top on MKA and had a substantially lower reported cost per task than the leader, but the benchmark’s private tasks and Google’s role in its development limit how broadly that result can be applied.
Key Questions
What did Gemini 4 Carbon score on MKA?
Carbon scored 54.3% on the MKA benchmark. Claude Fable 5.2 scored 56.6%, and Claude Opus 5.5 scored 53.0%; the source account says the top three results are within the margin of error.
How much does Carbon cost on MKA compared with Fable 5.2?
Google DeepMind’s report puts Carbon at $5.63 per MKA task and Claude Fable 5.2 at $35.28. These are benchmark-specific reported costs, not guaranteed prices for other tasks or deployments.
Who can use Gemini 4 Carbon now?
As of its Nov. 10 opening, Carbon is available to paying developers through the Gemini API and Google AI Studio, and to businesses through Gemini for business. Access in the Gemini app and AI Mode in Search is limited to Google AI Ultra subscribers.
Is Carbon the top model across all benchmarks?
No. Carbon ranked second on MKA, and Google’s model card shows it leading 15 of 17 tests selected by Google while losing two. Artificial Analysis’s current Intelligence Index also places Claude Fable 5.2 above Carbon.
How reliable is the MKA comparison?
The published results provide a comparison across 200 tasks, but Google co-developed MKA and the tasks are private, limiting independent replication. The close scores and stated margin of error also make a firm performance ranking difficult.
Source: ThorstenMeyerAI.com
