Gemini 4 Carbon
Multi-file debugging, architecture changes, technical research and analysis that combines charts with long explanations.
Chat with Gemini 4 Carbon here. Every message is a real request, billed to your balance at the same price as the API.
Usage-based pricing with separate rates for input, cached input and output. Thinking tokens are included in output usage.
| Model | Specifications and pricing | beatapi |
|---|---|---|
| Gemini 4 Carbon | Gemini 4 CarbonInput · per 1M tokens | $0.60 / 1M tokens |
| Gemini 4 Carbon | Gemini 4 CarbonCached input · per 1M tokens | $0.03 / 1M tokens |
| Gemini 4 Carbon | Gemini 4 CarbonOutput · per 1M tokens | $3.00 / 1M tokens |
Use Carbon for work that needs a deeper investigation across several sources or steps. Use Argon for quick, frequent interactions. Both fit the same Gemini Chat Completions integration, so you can choose by task within one application.
Multi-file debugging, architecture changes, technical research and analysis that combines charts with long explanations.
Everyday code questions, small feature changes, content organization and interactive assistance that benefits from quicker replies.
Gemini 4 ArgonCarbon works across code, technical text and images. Use it to explain how a problem happens, compare possible solutions and turn the analysis into a code change, a report or an action plan.
Combine the relevant code, error logs and failing tests to reconstruct the path from an operation to its failure. Carbon can return a root-cause analysis, focused code changes and regression tests covering the affected behavior.
Work from known conditions, measurements and equations to compare explanations and develop a reasoned answer. This is useful for interpreting experiments, checking a calculation or examining the trade-offs in a design.
Read a chart, screenshot or diagram alongside the question and supporting text. Carbon can describe trends, compare figures and explain structural relationships, then bring those observations into a report or decision brief.
Break a goal into steps and update the plan as new information arrives. Carbon can summarize progress, connect tool results with the original question and propose the next tool call or action for your application.
| Benchmark | Gemini 4 Carbon | Gemini 4 Argon | GPT-6 Astra | Claude Opus 5.5 | Claude Fable 5.1 |
|---|---|---|---|---|---|
| AA Index v5.0 | 59 | 50 | 57 | 58 | 49 |
| Vals Index v2.2 | 70.4% | 69.2% | 65.0% | 70.2% | — |
| SWE-sweep | 9.6% | 4.3% | 9.5% | 3.6% | 2.9% |
| FrontierSWE 2 | 61.2% | 55.0% | 65.5% | 62.3% | 56.3% |
| Terminal-Bench 4.0 (AA) | 65.2% | 57.1% | 59.1% | 59.6% | 52.0% |
| Agents' Last Exam v1.1 (Score) | 77.4% | 69.0% | 73.1% | 71.6% | — |
| OSWorld-2.1 (Offline subset) | 81.5% | 74.7% | 80.9% | 87.2% | 76.8% |
| AutomationBench (Zapier) | 53.9% | 51.3% | 41.4% | 42.5% | 31.4% |
| Vending-Bench 2 (Bank balance) | $16,484 | $13,718 | $15,515 | $9,235 | $5,422 |
| Chartography | 72.5% | 71.6% | 71.0% | 66.3% | 46.2% |
| Blueprint-Bench 2 (Similarity score) | 58.6% | 54.4% | 49.7% | 51.2% | 41.9% |
| GraphWalks (256k to 1M, BFS F1) | 96.2% | 84.2% | 71.8% | 66.8% | 65.0% |
| RiemannBench | 80.0% | 76.0% | 72.0% | 69.6% | 65.6% |
| Terminal-Bench-Science 0.1 (AA) | 64.8% | 52.9% | 63.3% | 59.0% | 43.3% |
| CWE-bench | 71.0% | 68.0% | 68.0% | 67.0% | 58.0% |
| ExploitGym | 54.8% | 42.0% | 42.4% | 34.5% | — |
| SEC-bench Pro | 88.2% | 83.4% | 85.4% | — | — |
October 2026 evaluation. Carbon scores 96.2% on GraphWalks for following relationships through long material, compared with Argon's 84.2%. On Terminal-Bench 4.0 engineering tasks, the scores are 65.2% and 57.1%. These results make Carbon worth testing for work such as connecting clues across technical documents and investigating bugs that span several files.
When retries or callbacks cause a payment problem, combine transaction-handling code with sanitized logs to reconstruct the sequence. Carbon can identify duplicate-processing paths and prepare a change plan with tests for the rule that each order is processed once.
For a service split, interface upgrade or database change, compare the current design with the target. Produce an implementation order, a map of affected modules and a compatibility checklist that can guide development and release.
For a report containing charts, screenshots and written conclusions, compare the figures and supporting passages together. Organize the result into conclusions with evidence, numerical discrepancies and questions to follow up, linked to the relevant source material.
Bring together measurements, experimental conditions and candidate explanations. Carbon can work through calculations, compare which explanations fit the observations and propose follow-up measurements, producing a research note that connects evidence to the next experiment.
Connect from your server with a BeatAPI key, choose Carbon, then read the returned answer and usage. Keep the OpenAI SDK and request format already used by your application.
Sign up or sign in to open the Dashboard. Check your balance and open the credits page if you need to add funds.
Open DashboardUse Carbon for technical work where the answer depends on following connections across code, documents and images, then turning that analysis into a fix, a report or a plan.
Connect code, specifications and technical records around the same question to investigate dependencies and trace a problem across documents. In the 256K–1M-token GraphWalks test, Carbon scores 96.2%, compared with Argon's 84.2%.
Bring investigation, code changes and test feedback into one continuing task, whether you are fixing a difficult bug or migrating several modules. Carbon scores 65.2% on Terminal-Bench 4.0, compared with Argon's 57.1%.
Combine a document's explanation with the figures and relationships shown in its charts to review reports, interpret diagrams and develop technical findings. On MKA's visual knowledge tasks, Carbon scores 54.3%, close to Fable 5.2's 56.6%.
Keep one BeatAPI integration and the OpenAI SDK. Route in-depth analysis to Carbon and everyday interactions to Argon, with the model name selected by your application instead of rebuilding the connection for each model.
| Benchmark | Gemini 4 Carbon | Claude Fable 5.2 | Claude Opus 5.5 | GPT-6 Astra |
|---|---|---|---|---|
| Quality (Climb WSR) | 54.3% | 56.6% | 53.0% | 50.2% |
| Cost / task (USD) | $5.63 | $35.28 | $10.29 | $6.15 |
November 2026 evaluation. MKA tests the ability to read visual material and turn it into useful work. Carbon scores 54.3%, close to Fable 5.2's 56.6%, at $5.63 per task versus $35.28. For report checking, diagram analysis and research that combines text with charts, compare Carbon's answer quality alongside its cost. Chart costs use standard model list prices.
Gemini 4 Carbon is Google's deep reasoning model for complex debugging, technical research and analysis that combines text with images. BeatAPI's OpenAI-compatible API lets you add these capabilities to developer tools, research assistants and report analysis applications, returning suggested fixes, findings and implementation plans.
Choose Carbon for difficult bugs spanning several files, architecture changes and research that requires comparing multiple sources. Keep Argon for everyday code questions, small changes and quick interactions. Their current BeatAPI token rates are the same, but Carbon generally spends more tokens thinking, increasing task cost and wait time. Your application can select the model by task while keeping the same request format.
The model has a shared 1,048,576-token window for your material, its thinking and the final answer. The default output budget is 65,536 tokens including thinking, not just visible text. The catalogue's 1,000,000-token maximum output involves long-output continuation; it is not the answer length to expect from one Chat Completions call. Leave space for thinking and the answer when preparing long inputs. See the API tab for integration parameters.
Output usage includes the model's thinking before it answers, not just the text you see. A difficult question can require substantial thinking even when the final answer is short. The response's usage fields report total consumption. Cache hits are billed at the cached-input rate; other input and output use the rates in the price table. Set max_tokens with room for both thinking and the answer.
Carbon can analyze images and respond with text. Send a chart, screenshot or diagram with supporting context to extract figures, explain structures or compare details. For example, send a chart with the report passage discussing it to check whether the figures support the conclusion. Choose an image model when you need to generate or edit pictures.
Have your application read the relevant files and send the code or report text with file names and section headings. Add charts as images. Explain the problem, what you have already found and whether you need a fix plan, comparison table or issue list. For debugging, include related modules, logs and failing tests so the analysis addresses the problem rather than simply summarizing the material.
Keep the OpenAI SDK, set the base URL to https://direct.beatapi.io/v1, supply a BeatAPI key and set model to gemini-4-carbon. Send messages through Chat Completions. The API tab includes request and response examples. Use the direct host for long-running tasks and allow enough time in your client's timeout settings.
Carbon can select a tool and generate its arguments; your application performs the action. Declare allowed functions in tools. When you receive tool_calls, have your server check permissions, run the operation and return the result with the matching tool_call_id. This connects code execution, database queries or internal APIs to the task while keeping access under your application's control.
Yes. Set stream: true to receive results incrementally. Carbon thinks before answering, and harder questions can take longer; streaming does not skip this thinking stage. Your application can show a processing state and read answer text from delta.content. Use direct.beatapi.io for long-running tasks and set an appropriate client timeout.
Place reusable instructions, code or documents at the start of the request, followed by the current question, and keep the shared portion stable. Input that hits the cache uses the cached-input rate in the price table. Check usage.prompt_tokens_details.cached_tokens to see how much was cached. Cache hits vary between requests, so billing follows the usage actually reported.
A regular chat request does not automatically perform a web search. For current research, have your application search or read pages, then send the relevant text, links and dates to Carbon for analysis. You can also expose search as a tool for the model to call during a task. This ties the findings to the material retrieved for that request.
No Google subscription is needed. Use a BeatAPI account, API key and funded balance. API usage is separate from subscriptions in the Gemini app; signup credit does not apply to Carbon. Once signed in with a funded account, you can try the same model in this page's Playground at API rates.
Connect Gemini 4 Carbon with other production-ready APIs in the BeatAPI catalog.
Guides and practical notes to help you get more from Gemini 4 Carbon.