Gemini 4 Carbon: what has been claimed about pricing, context, and supported formats

Gemini 4 Carbon is described in the available information as a multimodal AI model for use through an API and OpenRouter. The official launch has not yet been confirmed. All specifications, prices, and test results below refer to the claims made by the author of the original material and do not replace information from the developer.

What is known about the launch of Gemini 4 Carbon

There is still no official confirmation.

The claim that Gemini 4 Carbon was released “yesterday” remains a rumor. The provided materials contain no official press release, Google documentation page, model catalog entry, or other confirmation from the developer that would make the release date an established fact.

The distinction is significant for those choosing a model for a workflow. A name in a review, aggregator interface, or user test only indicates that such a claim exists or that a specific route is available from a provider. It does not prove that the model has received public stable status, identical access conditions for all users, or guaranteed support from Google.

As of the preparation of this analysis, the only confirmed points are statements recorded in the original description as claims by the author. The availability of Gemini 4 Carbon, the exact API model name, the official provider, pricing terms, and the duration of the discount need to be checked separately before integration, since publicly available materials do not provide reliable confirmation for them.

Context of 1 million tokens

The author claims a context window of 1 million tokens for Gemini 4 Carbon. If the developer confirms this specification in the documentation, the model will be able to process large volumes of text, code, and multimedia materials within a single task; however, the window size alone does not guarantee consistently high response quality throughout a long request.

A large context is useful when analyzing large repositories, multipage technical documentation, lengthy transcripts, file collections, and scenarios where early conditions must be compared with later changes. A team could provide the model with several related documents, ask it to find contradictions, and then generate a summary without breaking the task into a large number of separate conversations.

The price depends not only on the size of the available window but also on the actual number of tokens processed, the billing direction, caching, and repeated runs. A request that technically fits within 1 million tokens may still be expensive or slow, especially with a high reasoning level and multiple task continuations.

Before using a large context, check how the provider counts tokens, whether the limit applies to all types of input data, whether it also applies to output, and whether the stated capacity remains available when working through OpenRouter. The description does not confirm that 1 million tokens is available in all modes and from every provider.

Claimed cost of Gemini 4 Carbon

The original material mentions a temporary discount of 50%. The author gives the following figures: 2 dollars for input, 10 dollars for output, and about 0,10 dollars for cached tokens. These amounts should be regarded as claimed terms for test access, not as a confirmed official price list.

Without a unit of measurement, these figures cannot be used for an accurate budget calculation. Language-model rates are usually specified for a particular volume of tokens, so the wording “2 dollars for input” is insufficient by itself for comparison with other models. It is necessary to determine whether the aggregator adds its own commission, whether the rate varies by provider, and whether the discount applies to multimodal data.

Caching reduces costs when the same large context is transmitted repeatedly, such as project documentation or a set of instructions. The stated cache price does not mean that every repeated request will be billed at the minimum rate: the rules may depend on the cache lifetime, prefix matching, and the way the request is routed.

What happens if these amounts are used without verification? The integration budget will be inaccurate, and the discount may end before the product launches. For a production system, it is best to first run a small series of requests, compare the actual bill in the provider's dashboard, and only then scale the workload.

What data formats are claimed

The description names five types of input data for Gemini 4 Carbon: audio, files, images, text, and video. The list indicates the model's claimed multimodality but does not disclose the technical details of processing each format.

• Audio. Supported containers, codecs, recording duration, and the ability to separate voices need to be checked.
• Files. It is important to clarify the permitted extensions, maximum size, upload method, and whether the document structure is preserved.
• Images. Resolution, the number of images in a single request, and the availability of analysis for tables, diagrams, and small text should be checked.
• Text. The character or token limit, rules for working with code, and behavior when languages are mixed need to be determined.
• Video. The video duration, frame-extraction rate, audio-track processing, and the ability to refer to specific moments in the recording are relevant.

Format support does not equal a guarantee of accurately understanding the content. The model may accept a video file, but the quality of analysis will depend on the duration, scene dynamics, resolution, audio, and the way the material is transmitted through a specific API route. Multimodal capabilities should be tested with separate scenarios rather than one universal request that would not reveal differences between formats.

What the author's tests showed

The author conducted tests through OpenRouter and used the highest available reasoning level, High. The results describe a specific test scenario, intermediary configuration, and the author's subjective assessments. They should not be turned into a universal ranking of Gemini 4 Carbon or proof of superiority over other models.

In the benchmark mentioned, Carbon, according to the author, scored higher than Fable 5.1, Opus 5, Astra, and Argon. Fable 5.2 was not included in the comparison because, as claimed, the model appeared shortly before the test. The author's assumption that Fable 5.2 would have produced a result no better than or roughly at the same level remains an assumption and is not confirmed by the comparison conducted.

The growing-flower test took the author 29 minutes 48 seconds and cost 1,43 dollars. The author rated the visual result and animation as comparable to Fable 5.2, noting a cinematic cartoon style and additional droplets. The possible error in revealing the seed was also identified by the author as a personal observation, not as a measured defect confirmed by independent verification.

In the orc scenario, generation took 20 minutes 44 seconds and cost approximately 0,03 dollars. The author noted a correct wind-up and ax strike but pointed out character flaws: the fangs resembled horns, the fingers were only partially rendered, the hairstyle looked like a cone, and there was a ring-shaped intersection in the body. In the author's assessment, the result was good and at least comparable to Opus, although it may not have reached Fable's level.

The aquarium took about two hours to create because of several failures accompanied by a message that the servers were unavailable, and cost 6,13 dollars. The scene included rocks, a crab with moving claws, snails, fish, plants, a castle, an aerator, and a dangling wire, but the author found errors in the placement of shells, the transparency of the fish, the crab's coloring, and the snail's interaction with the rocks.

According to the author's observation, generating the platformer effectively took about two hours, although the interface displayed 38 seconds. The author did not consider the displayed time reliable because they pressed Continue repeatedly. The robot game received traps, shooting, parkour, momentum, and slow motion while carrying a battery; the author called its similarity to the previously tested Fable game noticeable and rated the overall quality as no worse than Fable, although in the author's opinion Fable's world was more polished.

The design-studio website was created in 22 minutes at a cost of 1,18 dollars. It included an interactive three-dimensional object that responded to the mouse, and the author rated the combination of pink-blue and yellowish colors positively. Responsiveness was incomplete: some elements adapted to the screen, while others did not.

How OpenRouter and High mode affect the result

In this description, OpenRouter acts as an intermediary through which the author ran the tests. This access method adds the work of the router, selected provider, queue, and task-continuation interface to the model evaluation. If requests pass through different nodes, the same model name does not guarantee the same speed, availability, or final cost.

High mode means the highest reasoning level available to the author in the test configuration. It may increase execution time and token usage, but the provided materials do not allow an exact relationship to be derived between reasoning level, result quality, and cost. The author themselves associated some failures and lengthy execution with OpenRouter, although Google AI Studio and Vertex appeared among the providers; this observation does not prove the cause of the problems.

For model comparisons, the same prompts, source files, time limits, reasoning mode, and continuation rules need to be preserved. Otherwise, the result reflects not only the model's capabilities but also the influence of the route, number of iterations, manual interventions, and available budget.

What to check before integration

1. Confirm that Gemini 4 Carbon exists in the developer's official catalog or with the selected API provider.
2. Verify the exact model name, version stability, and date of the latest pricing update.
3. Clarify the billing units for input, output, and cached tokens.
4. Check whether the 50% discount is active, when it ends, and whether it applies to all data types.
5. Test audio, files, images, text, and video with separate requests.
6. Measure latency, number of failures, token usage, and cost on your own scenarios.
7. Set a budget limit and configure error handling, retries, and termination of lengthy tasks.
8. Check which provider actually executes the request and whether user data is transmitted to it.

For a prototype, the claimed specifications look interesting: a large context, multimodal input, and the ability to run lengthy tasks through an intermediary. This is not enough for a production integration. First, confirmed model details, a stable route, transparent pricing, and repeatable tests on the data the system will process after launch are needed.

Summary

In the available description, Gemini 4 Carbon is associated with a context of 1 million tokens, input for audio, files, images, text, and video, as well as claimed rates of 2 dollars for input, 10 dollars for output, and about 0,10 dollars for caching. The release date, official model status, and validity of the discount have not been confirmed, so they should be described as a rumor or unconfirmed claim.

The author’s tests show interesting results in generating 3D scenes, games, and websites, but remain a single experiment conducted through OpenRouter in High mode. The author’s overall assessment is as follows: in some tasks, the system approached Fable 5.2; in others, it fell short of it; and overall, it was roughly comparable to Opus 5. This is the author’s opinion, not an official result from the developer or a universal conclusion about the quality of Gemini 4 Carbon.