1M Unified Context
Budget prompts, thinking, function calls, and generated text inside one 1,048,576-token window instead of treating input and output as separate allowances.
curl --request POST "https://gptproto.com/v1/chat/completions" \
--header "Authorization: Bearer $GPTPROTO_API_KEY" \
--header "Content-Type: application/json" \
--data '{
"model": "gemini-4-carbon",
"messages": [
{
"role": "user",
"content": "Hello"
}
]
}'Chat, Coding-Agenten und Dokumentenarbeit. Preise pro 1 Mio. Tokens – Eingabe, gecachte Eingabe und Ausgabe werden separat abgerechnet. GPTProto liegt 30% unter den offiziellen Preisen.
| Szenario | Google Liste | OpenRouter | GPTProto | Du sparst / Monat |
|---|---|---|---|---|
| Persönlich10 Mio. Tokens / Monat (4.8 Mio. im Cache) | $53.76 | $56.72 | $37.63 | −$16.13≈ $193.54 / Jahr |
| Team100 Mio. Tokens / Monat (48 Mio. im Cache) | $537.60 | $567.17 | $376.32 | −$161.28≈ $1935.36 / Jahr |
| Business500 Mio. Tokens / Monat (240 Mio. im Cache) | $2688.00 | $2835.84 | $1881.60 | −$806.40≈ $9676.80 / Jahr |
Gemini 4 Carbon is Google's frontier model for projects that require repeated planning, tool use, verification, and revision. It combines a unified 1,048,576-token context with configurable reasoning and multimodal input. On GPTProto, developers can test Carbon alongside other supported frontier models without managing a separate key and balance for every provider.
1M Unified Context
Budget prompts, thinking, function calls, and generated text inside one 1,048,576-token window instead of treating input and output as separate allowances.
Long Decodes for Large Deliverables
Carbon can continue beyond its normal synchronous output limit through supported long-decode flows, leaving more room for large code changes, proofs, and technical reports.
Low, Medium, or High Thinking
Choose low, medium, or high reasoning depth by task. Medium is Google's default; higher levels can use more time and tokens on difficult work.
Multimodal Tools and Structured Results
Analyze text, images, and documents on GPTProto, then return text, structured data, or tool calls for agent workflows that your application can validate.
Gemini 4 Carbon is a Google DeepMind reasoning model released on November 10, 2026. Google positions it above Gemini 4 Argon for tasks that test the limits of a model's intelligence, including complex multistep coding, difficult debugging, open-ended scientific research, cybersecurity analysis, and long-horizon knowledge work. Carbon is based on Argon but uses more reasoning tokens to explore alternatives and run extended verification loops.
The underlying model accepts text, images, audio, and video and returns text. The current GPTProto model page exposes text, image, and document input with text output. This distinction matters: applications should use only the modalities and request fields shown in GPTProto's live API Usage tab, even when Google's native API supports additional inputs.
Unlike models with separate input and output ceilings, Carbon uses a shared 1,048,576-token budget. Input, reasoning, function calls, and the final answer all count toward that limit. A larger prompt therefore leaves less room for thinking and generated output. This is particularly important for repository-scale coding and long agent runs, where tool results can expand the conversation quickly.
GPTProto provides third-party access to the Google Gemini 4 Carbon API through an OpenAI-compatible chat endpoint. The same account can also call other Google models listed in the Google API catalog, making it easier to compare Carbon with faster or more specialized routes before assigning production traffic.
| Specification | Gemini 4 Carbon on GPTProto |
|---|---|
| Developer | Google DeepMind |
| Release date | November 10, 2026 |
| GPTProto model string | gemini-4-carbon |
| Current GPTProto input / output | Text, image, and document to text |
| Context model | Unified input-and-output context |
| Total context limit | 1,048,576 tokens |
| Standard synchronous output | Up to 65,536 tokens in supported Google API flows; direct non-streaming HTTP requests are limited to 32,768 tokens |
| Long-decode ceiling | Up to the unused portion of the shared 1,048,576-token window through supported continuation or background flows |
| Thinking levels | LOW, MEDIUM, and HIGH; Google lists MEDIUM as the default |
| Agent features | Function calling, streamed tool arguments, structured output, and multimodal analysis |
| Availability | Google API live; GPTProto route live |
Implementation note: Output limits, continuation behavior, thinking controls, and provider-native tools can differ across API formats. Follow the live GPTProto request schema rather than copying Google-specific fields into an OpenAI-compatible request without testing them.
Repository-scale software engineering. Give an agent source files, interfaces, tests, logs, and acceptance criteria. Carbon fits work that requires tracing dependencies, running tools, inspecting failures, and revising an implementation rather than returning a first-pass code sample.
Long-horizon agentic work. Use the model for loops that plan, call tools, evaluate results, and recover from failed steps. Keep permissions, state, timeouts, and acceptance tests in the application.
Multimodal knowledge work. Combine documents, screenshots, charts, diagrams, and instructions for evidence synthesis, structured reporting, spreadsheet planning, or visual analysis. Confirm supported formats before sending a large batch.
Scientific, analytical, and defensive review. Carbon can compare hypotheses, check intermediate results, review code, and plan remediation. Expert review remains necessary. The restricted gemini-4-carbon-cyber-permissive variant is separate and is not documented as a GPTProto route.
Google DeepMind's October 2026 model card reports the following results for Carbon and Gemini 4 Argon. These values are useful for understanding the intended upgrade, but they were not independently reproduced by GPTProto and should not replace tests on your own prompts, tools, token budgets, and acceptance criteria.
| Evaluation | Gemini 4 Carbon | Gemini 4 Argon | What It Tests |
|---|---|---|---|
| SWE-sweep | 9.6% | 4.3% | Difficult agentic software engineering |
| Terminal-Bench 4.0 | 65.2% | 57.1% | Tool-driven terminal tasks |
| AutomationBench | 53.9% | 51.3% | Multistep knowledge-work automation |
| GraphWalks, 256K-1M | 96.2% | 84.2% | Long-context graph traversal |
| RiemannBench | 80.0% | 76.0% | Advanced mathematical reasoning |
The launch model card compares Carbon with earlier GPT and Claude versions, not GPT-6.2 Astra or Claude Fable 5.2. Do not relabel those older scores as results for the newer models.
No frontier model wins every workload. The practical differences are reasoning controls, output behavior, provider-specific tools, and how much latency a task can tolerate. Use the same repository, tool schema, prompt, and pass/fail tests when comparing routes.
| Alternative | Key Difference from Gemini 4 Carbon | Choose Carbon When | Choose the Alternative When |
|---|---|---|---|
| GPT-6.2 Astra | Astra has a 1.05M context window, a 128K maximum output, and OpenAI reasoning levels from low through max. Carbon uses a unified 1,048,576-token budget and supports continuation-based long decodes in Google's native flows. | The task needs Google's multimodal reasoning, deep scientific work, or extended verification over a shared context budget. | Your system is built around OpenAI's Responses API, computer use, or OpenAI-specific agent features. |
| Claude Fable 5.2 | Fable provides a 1M context window, up to 128K output, and always-on adaptive thinking with five effort levels. | You want Carbon's Google ecosystem, three explicit thinking levels, and reported strength in long-context, science, and automation evaluations. | You prefer Anthropic's tooling and need its long-running technical-work behavior or Claude-specific integration features. |
| Gemini 4 Neon | Google describes Neon as a Carbon-derived model that balances depth, speed, and cost for complex high-frequency workflows. Carbon remains the option for the hardest long-horizon projects. | Accuracy and extended verification matter more than response speed or token efficiency. | The workflow runs frequently and needs a better depth-to-latency balance. Use the Google model catalog until a verified GPTProto Neon route is available. |
| Kimi K3.1 Pro | Kimi also uses a shared 1M-token window, but its reasoning stays on and uses low, high, or max effort. Its GPTProto route is positioned for long coding, visual-to-code, and agent work. | You need Google's multimodal knowledge-work and scientific-reasoning stack or want to reproduce Carbon's reported evaluation strengths. | You want Moonshot's Kimi workflow, preserved-thinking behavior, or a second long-context model for canary testing. |
For most routine Google-model software engineering and knowledge work, Gemini 4 Argon remains the more natural baseline. Use Carbon when the job justifies deeper exploration, longer verification loops, and potentially higher latency.
Budget the shared context. Leave room for thinking, tool calls, and the final response instead of filling the window with source material.
Confirm reasoning controls. Google's native API uses LOW, MEDIUM, and HIGH; older thinking_budget or MINIMAL settings can fail. Use only parameters documented by GPTProto.
Do not depend on sampling knobs. Gemini 4 ignores custom temperature, top_p, and top_k values. Use thinking level and structured output instead.
Test long decodes. Continuation tokens and background execution are not guaranteed through every compatible endpoint. Run an output-length test first.
Validate tools and retries. Restrict permissions, verify arguments, require acceptance tests, and plan for occasional slow or timed-out requests.
Choose Gemini 4 Carbon when a task is difficult enough to benefit from deeper reasoning and repeated verification: cross-repository changes, open-ended debugging, scientific analysis, dense multimodal evidence, or agents that must work through many dependent steps. It is a poor default for simple extraction, classification, or short chat if a faster model can meet the same acceptance test.
Compare representative tasks on success rate, tool-call accuracy, wall-clock time, output-token use, and recovery from failed steps. GPTProto's shared key lets the application switch among supported models without opening another provider account.
Anleitungen, Vergleiche und Updates zu diesem Modell.
Alle Artikel
What is GPT-6.2 Astra? See its real-world uses, API pricing, early reviews, Astra vs Sol and Fable 5.2, plus 20% off on GPTProto.

We tested 4 GPT models for character chat, romance, and RPGs. Compare dialogue, speed, costs, and Kimi results to find the right fit for your roleplay.

What is Qwen 4 Max (qwen4-max-preview)? Explore its 1M context window, 393K output cap, Token Plan rules, hands-on tests, and pay-as-you-go API options.

Compare 11 all-in-one AI API platforms for language, image, video, and audio workflows. Check current modality coverage, billing units, free-access scope, and API formats.