🔍 Read the full analysis: Could Gemini 4 Carbon Be Worth A Look At One-Sixth The Cost? on ThorstenMeyerAI.com
TL;DR
Google opened Gemini 4 Carbon to paying developers, businesses and AI Ultra subscribers on November 10, alongside a model card. On a new benchmark co-developed by Google DeepMind and Artificial Analysis, Carbon scored 54.3%, just behind Claude Fable 5.2 at 56.6%, while Google reported a lower cost per task: $5.63 versus $35.28.
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 visual documents. Carbon ranked second at 54.3%, behind Claude Fable 5.2 at 56.6%, but Google DeepMind reported that Carbon cost $5.63 per benchmark task compared with $35.28 for Fable 5.2.
The new test, called MKA, consists of 200 expert-built tasks spanning 10 sectors. It tests agents on material such as blueprints, scans and charts, rather than relying only on plain-text prompts. Claude Opus 5.5 scored 53.0%, putting the three leading models within a narrow range; the source report says their results are within the margin of error.
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 limited to Google AI Ultra subscribers. Google set introductory API pricing at $2 per million input tokens and $10 per million output tokens through November 30. From December 1, those rates rise to $4 and $20 per million tokens.
Google’s model card reports that Carbon led on 15 of 17 benchmarks Google selected, against four listed competitors. It lost to GPT-6 Astra on FrontierSWE 2 and Claude Opus 5.5 on OSWorld-2.1. The card’s results are dated October 2026 and do not include Claude Fable 5.2, which was released on November 9. The figures therefore describe Google’s chosen comparison set, not the newer MKA ranking.
Could Gemini 4 Carbon Be Worth a Look? A lower benchmark cost, with trade-offs
Gemini 4 Carbon scored close to Claude Fable 5.2 on a new visual-document benchmark, while Google reported a much lower cost per task. The result may interest businesses with document-heavy workloads, but does not predict every customer’s cost or needs.
A close benchmark result
MKA tests AI agents on visual documents such as blueprints, scans and charts. Its leading scores are within the margin of error, according to the source report.
Second on MKA
Its 54.3% score sits just below Claude Fable 5.2.
First on MKA
It scored 56.6%, but at a higher reported cost per task.
Close behind
Claude Opus 5.5 scored 53.0%.
Where access is available
The November 10 launch opened paid access, with consumer app access more limited.
- Paid Gemini API
- Google AI Studio
- Gemini for business
- Gemini app requires Google AI Ultra
- AI Mode in Search requires Google AI Ultra
Pricing changes in December
Introductory API rates end November 30; Google lists higher rates from December 1.
| API rate | Through November 30 | From December 1 |
|---|---|---|
| Input, per million tokens | $2 | $4 |
| Output, per million tokens | $10 | $20 |
A specialist with practical trade-offs
Google positions Carbon for demanding, multi-step work; Argon remains its choice for most software engineering and knowledge work.
Carbon handles text, images, audio and video input, and produces text.
Artificial Analysis reports 25.6 minutes per index task for Carbon, compared with a 12.6-minute median for comparable models.
Prompt and output length, settings, retries and token use can change customer costs.
Cost Versus Benchmark Performance
For companies evaluating AI agents, the MKA result puts a practical trade-off in focus: Carbon scored slightly below Fable 5.2, while the reported task cost was much lower. If that cost difference holds for comparable work, it could matter to organizations running many document-heavy tasks, where per-task expenses can accumulate.
The result does not establish that Carbon is the better-value choice for every workload. MKA covers a defined set of 200 tasks, and the reported costs are benchmark-specific. A company’s actual bill can depend on prompt length, output length, model settings and how often a task needs to be retried. Carbon also takes longer on Artificial Analysis’s separate index: it reports 25.6 minutes per index task, compared with a 12.6-minute median for comparable models. Lower cost per task may not mean faster completion or a better fit for time-sensitive work.
For individual users, access is more restricted in Google’s consumer products than the headline availability might suggest. Gemini app and Search AI Mode access requires AI Ultra; the broader launch is aimed at paid developer and business services. The pricing promotion also has a fixed end date, so users comparing costs need to distinguish the introductory rates from the rates scheduled to apply in December.
Popular right now
Shark Rotator Lift-Away Upright Vacuum with DuoClean PowerFins, Wine Purple, LA702
Hisense E6SR Series 55-Inch 4K UHD Hi-QLED Roku Smart TV
Vitamix 5200 Blender with 64 oz. Container, Black
As an affiliate, we earn on qualifying purchases.
A Specialist Model With Trade-Offs
Google describes Carbon as its top-tier model for demanding tasks, including complex multi-step coding, debugging and open-ended scientific research. It is built on Gemini 4 Argon and uses more thinking tokens, according to Google, to explore possibilities and run extended verification. Google says Argon remains the better choice for most software engineering and knowledge work, positioning Carbon as a specialist rather than a universal replacement.
The model accepts text, images, audio and video, but produces text. Its context window supports up to 1 million tokens in total across input and output, and Google lists a March 2026 knowledge cutoff, with a warning that information may be less current in some domains. Developers can choose low, medium or high thinking levels; medium is the default. Google also identifies hallucinations, occasional slowness or timeouts, and higher token use at stronger thinking settings as limitations.
Carbon was first made available to a small group of cyber defenders through Google’s Fairwind Program about two weeks before the wider paid launch; Artificial Analysis lists October 28 as its release date. Its stronger security features remain limited to verified defenders, according to the source report. The November 10 launch thus broadened paid access, but did not make every feature available to every user.
„“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 co-developed by Google DeepMind, and the source report notes that Google’s own model performs well on it. The tasks are private, so outside researchers cannot inspect them or independently reproduce the scores. The benchmark’s result is therefore useful as a reported comparison, but its design and the unpublished tasks limit outside scrutiny.
The reported cost figures also need careful interpretation. Google DeepMind’s report gives Carbon’s MKA cost as $5.63 per task and Fable 5.2’s as $35.28, but the source material does not provide enough detail here to establish how those figures would translate to a particular customer’s workload. The MKA scores are close, and the source says the top three are within the margin of error; it does not establish a statistically decisive performance gap.
Other comparisons use different methods. Artificial Analysis’s Intelligence Index v5.0.1 gives Carbon at its high setting a score of 60, versus Fable 5.2 at 62 on high and 65 at its maximum setting. That index combines nine tests and is not the same evaluation as MKA. It is not yet clear how Carbon will perform across a broader range of independent, public tasks or how its speed and token use will affect real-world costs.
Pricing and Independent Testing
The immediate date to watch is December 1, 2026, when Google’s listed API rates are scheduled to rise from the introductory $2/$10 per million input and output tokens to $4/$20. Developers considering Carbon can compare the launch-period price with the standard rate, while businesses will need to test their own workloads rather than assume benchmark task costs match their usage.
Further independent evaluations could clarify whether Carbon’s MKA ranking carries over to other visual-document and agent tasks. For now, the available evidence shows a model that is accessible to paying users, ranks near the top on a new but co-developed benchmark, and has a lower reported MKA cost than the leading model. Its performance outside the published tests, practical speed, and customer-specific costs remain open questions.
Key Questions
Who can use Gemini 4 Carbon now?
Google opened Carbon to paying developers and businesses through its 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.
How did Carbon perform on the new 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 report says the leading results are within the margin of error.
Is Carbon one-sixth the price of Claude Fable 5.2?
On MKA, Google DeepMind reported a cost of $5.63 per Carbon task and $35.28 per Fable 5.2 task. That is roughly one-sixth on this benchmark, but it does not show that Carbon will cost one-sixth as much for every customer or type of work.
When does Carbon’s API price change?
The introductory rate runs through November 30, 2026: $2 per million input tokens and $10 per million output tokens. Google’s listed rates from December 1 are $4 and $20 per million tokens, respectively.
What are the main limitations known so far?
Google lists hallucinations, occasional slowness or timeouts, and higher token use, especially at higher thinking levels. Artificial Analysis also reports a 25.6-minute average per index task, though that figure comes from a separate evaluation and does not predict every workload.
Source: ThorstenMeyerAI.com
