Swapping the model string to gemini-4-carbon does not finish the request. On 10 November 2026 Google opened Gemini 4 Carbon on the paid Gemini API, in Google AI Studio, and for AI Ultra subscribers in Antigravity. Introductory price through 30 November is $2 per million input tokens and $10 per million output tokens, and output includes thinking tokens. From 1 December those rates are $4 and $20. Implicit cache hits are 95 percent off ($0.10, then $0.20). Carbon is not on the free tier.
This page is dated 13 November 2026 from the 10 November developer guide and the What's new page. Three days ago we covered Fable 5.2 effort and usage. Carbon's controls live on the Interactions API: generation_config.thinking_level, background, and usage.
What opened on 10 November
Google calls Carbon its strongest reasoning model for long-horizon engineering, dense multimodal files, and scientific work. The model id is gemini-4-carbon. Default thinking level is medium. Input, thinking, and output share one 1,048,576-token window. A synchronous response still stops at 65,536 tokens unless you opt into long decode.
gemini-4-argon shares that window, the same thinking_level values, built-in tools, and the same introductory token price. Carbon spends more thinking tokens per turn. Keep Argon on interactive coding. Move Carbon onto root-cause work, architecture, and verification loops.
| Window | Input / 1M | Output, including thinking / 1M | Cached input / 1M |
|---|---|---|---|
| Through 30 November 2026 | $2 | $10 | $0.10 |
| From 1 December 2026 | $4 | $20 | $0.20 |
Request JSON and thinking_level
POST https://generativelanguage.googleapis.com/v1beta/interactions with x-goog-api-key and Content-Type: application/json. Depth is generation_config.thinking_level, not a numeric budget.
{
"model": "gemini-4-carbon",
"input": "Analyze this payment pipeline for race conditions during retries and rewrite the locks.",
"generation_config": {
"thinking_level": "medium"
}
}
| thinking_level | What the Carbon guide says to use it for |
|---|---|
| low | Routing, classification, and targeted extraction |
| medium | Default. Most coding and agent turns |
| high | Root-cause debugging, formal math, architecture |
Omit thinking_level and the server uses medium. The Carbon guide documents low, medium, and high. The shared Interactions enum also lists minimal and xhigh for other models; do not send those on Carbon until the Carbon page lists them. thinking_budget is deprecated and returns a warning. max_output_tokens caps thinking plus visible text together, so a low cap can cut the answer off before the model finishes reasoning.
background and long decode
To generate past 65,536 tokens, set top-level background to true and set max_output_tokens above 65536. The ceiling is 1,048,576 minus the input. background: true without max_output_tokens still stops at 64K.
{
"model": "gemini-4-carbon",
"input": "Write a complete proof of the prime number theorem.",
"background": true,
"generation_config": {
"thinking_level": "high",
"max_output_tokens": 262144
}
}
The Carbon guide polls while status is in_progress or queued. The Interaction resource also uses completed, failed, cancelled, incomplete, requires_action, and requires_continuation. incomplete means the cap was hit. Over HTTP, read the id with GET /v1beta/interactions/{id}. The getting-started background sample sends header Api-Revision: 2026-05-20 on that GET.
Each long-decode hop appends another steps[] item with type: model_output. Join every text part. Reading only steps[0] drops the rest. previous_interaction_id continues history and does not replay generation_config, tools, or system_instruction; send thinking_level again. Default store is true (paid tier keeps interactions 55 days). store: false cannot be combined with background: true.
Fields to remove from older Gemini JSON
Moving from Gemini 3 or earlier, the What's new page and the developer guide tell you to change the JSON, not only the model id:
temperature,top_p, andtop_kare ignored. The developer guide says they return a deprecation warning.- Remove explicit
cached_contentandcaches.create. Implicit caching starts at 4,096 tokens and discounts cache hits by 95 percent. - Replace
thinking_budgetwiththinking_level. - If
max_output_tokensis under 32,768, raise it to at least 32,768. Lower the thinking level when you need fewer tokens. - Hard prompts can take minutes. Use background execution, or set the HTTP client timeout to at least 15 minutes.
Structured JSON for your own schema still belongs on response schema fields. See structured JSON with the Gemini API and what structured output is. Carbon does not turn a decision into probabilities; that is a different endpoint.
usage before the 1 December price
A finished interaction carries usage. In the published sample, 7 input + 20 output + 22 thought tokens equal total_tokens 49, so total_thought_tokens is not inside total_output_tokens.
{
"id": "v1_example",
"object": "interaction",
"model": "gemini-4-carbon",
"status": "completed",
"steps": [
{
"type": "model_output",
"content": [{ "type": "text", "text": "..." }]
}
],
"usage": {
"total_input_tokens": 7,
"total_cached_tokens": 0,
"total_output_tokens": 20,
"total_thought_tokens": 22,
"total_tool_use_tokens": 0,
"total_tokens": 49,
"input_tokens_by_modality": [{ "modality": "text", "tokens": 7 }]
}
}
Bill uncached input at the input rate, total_cached_tokens at 5 percent of that rate, and total_output_tokens plus total_thought_tokens at the output rate. Do not also add total_tokens. Grounding still bills search queries separately, and tool instructions count as processed tokens. Store the whole usage object so a 1 December price change can be applied to logs you already have.
Argon routing and agent_config
Express the split in config instead of one global model. Argon stays on the fast path at the same list price; Carbon takes the turns that spend thinking tokens.
{
"routes": [
{ "if": { "task": "interactive_coding" }, "model": "gemini-4-argon", "thinking_level": "medium" },
{ "if": { "task": "root_cause" }, "model": "gemini-4-carbon", "thinking_level": "high", "background": true }
]
}
Managed agents set the model inside agent_config. generation_config applies when model is set; agent_config is the alternative.
{
"agent": "antigravity-preview-09-2026",
"environment": "remote",
"input": "Audit the site for Core Web Vitals and return a scorecard.",
"agent_config": {
"type": "antigravity",
"model": "gemini-4-carbon"
}
}
JSON to verify before launch
- Request body has
gemini-4-carbon,thinking_levelin low/medium/high, and notemperature,top_p,top_k,thinking_budget, orcached_content. - Any call that must pass 64K tokens sets
background: trueandmax_output_tokensabove 65536. - The parser joins every
model_outputtext part and recordsstatus. - Billing uses input, cached input, output, and thought tokens, with the 1 December rates ready.
Check the JSON locally with the JSON validator, JSONPath ($.usage.total_thought_tokens), and JSON Diff against the previous request. Nothing is uploaded.
FAQ
Is gemini-4-carbon on the free tier?
No. The What's new page says Carbon is paid-tier only. Introductory $2/$10 pricing runs through 30 November 2026, then $4/$20.
Can I send thinking_level xhigh?
The Carbon guide documents low, medium, and high. minimal and xhigh appear on the shared Interactions enum for other models. Do not copy Claude effort values into this field.
Why does the answer stop at 64K tokens?
Synchronous responses stop at 65,536. Set background to true and max_output_tokens above 65536. background alone still stops at 64K. status incomplete means the cap was hit.
Should the bill use total_tokens?
No. Price uncached input at the input rate, total_cached_tokens at 5 percent, and total_output_tokens plus total_thought_tokens at the output rate. Adding total_tokens double-counts.
Does previous_interaction_id keep thinking_level?
No. Resend generation_config, tools, and system_instruction on the next interaction. History is what the id preserves.
Can I keep an explicit context cache?
No. Remove cached_content. Implicit caching applies from 4,096 tokens and discounts hits by 95 percent.
Summary
Carbon, opened 10 November 2026, changes how long the model thinks and what you pay after 30 November. Use gemini-4-carbon with thinking_level at low, medium, or high. Pair background: true with max_output_tokens above 65536 when the answer must pass 64K. Bill thought tokens separately from visible output, and drop sampling fields and explicit caches before the request goes out.