Introduction
Send text and images in the OpenAI chat format and read back the reply, function calls or JSON. Reasoning happens before the answer and draws on max_tokens, so set the cap well above the length you expect to read.
| Field | Value |
|---|---|
| Input | Text |
| Output | Text |
| Shared context | 1,048,576 tokens |
| max_tokens ceiling | 65536 |
| Inputs | text / images |
Key features & highlights
Reserved for the hard problems
Google places Carbon above Gemini 4 Argon for multi-step debugging and open-ended research, while noting that Argon still suits most everyday engineering work.
It spends tokens on checking itself
Carbon uses more tokens to reason and verify than lighter models, so expect a long hidden working phase before a short final answer.
One window for the whole problem
Logs, source files and notes can share a single 1,048,576-token context with the reasoning and the reply.
Request parameters
| Parameter | Required | Description |
|---|---|---|
messages | Yes | The Carbon conversation, oldest first. User messages can combine text parts with |
max_tokens | No | Carbon reply budget from 1 to 65536, default 4096. Reasoning is counted here: a 400-token cap on a long prompt used 397 tokens on reasoning and returned no text. |
reasoning_effort | No | Carbon thinking depth: low, medium or high, with medium as the default. Thinking cannot be turned off, and high uses the most tokens. |
stop | No | One to four stop strings of up to 32 characters each. Carbon's reply ends before the first match, and the match is not returned. at least 1 characters · at least 1 items · at most 4 items |
temperature | No | Accepted from 0 to 2 so OpenAI-style clients keep working. Google's documentation says Carbon ignores this setting, so it will not change the reply. |
top_p | No | Accepted from 0.01 to 1 so OpenAI-style clients keep working. Google's documentation says Carbon ignores this setting as well. |
tools | No | Up to 128 function definitions Carbon can request. Your application validates the arguments, runs the function and returns the result in a tool message. at most 128 items |
parallel_tool_calls | No | On by default, so Carbon can request several functions in one turn. Set it to false to receive one call at a time. Default: |
tool_choice | No | Sets Carbon's next move: auto, none, required, or one named function. |
response_format | No | Asks Carbon for a JSON object, or for JSON that follows a |
Pricing
Input $2 / Output $10 per 1 million tokens
Use cases
Chase a bug across several services
Paste the failing logs and the relevant code, then ask Carbon to rank likely causes and name the check that would rule each one out.
Test a scientific argument
Give Carbon a derivation or method section and ask where the reasoning depends on an assumption that has not been stated.
Prompt guide & pro tips
Leave plenty of room in
max_tokensThinking is drawn from the same budget as the answer. In our test, a 400-token cap on a long prompt spent 397 tokens on thinking and returned no text.
Save Carbon for work that needs it
Send routine rewrites and lookups to a lighter model, and bring Carbon in when a wrong answer would cost real time.
Gemini 4 Carbon is Google's strongest reasoning model for hard debugging, research and long agent runs.
google/gemini-4-carbon/chat — /v1/chat/completions
Set reasoning_effort to low, medium or high; Carbon starts at medium. Reasoning draws on max_tokens before any visible text appears.
If the cap runs out during reasoning, the call ends with finish_reason length and an empty reply.
User messages can carry image_url parts. Carbon can also request the functions you declare in tools and return JSON through response_format.
Google's safety systems review Carbon requests, so some prompts may be declined.
API quick start
curl --fail-with-body -sN https://api.spicyapi.ai/v1/chat/completions \
-H "Authorization: Bearer $SPICY_API_KEY" \
-H "Content-Type: application/json" \
-H "Idempotency-Key: $(uuidgen)" \
-d '{
"model": "google/gemini-4-carbon/chat",
"messages": [
{
"role": "user",
"content": "Summarise the following release notes into three bullet points."
}
],
"stream": true
}' \
| jq --unbuffered -Rrj 'select(startswith("data: ")) | .[6:] | select(. != "[DONE]") | fromjson | if .error then "\n[error] \(.error.message)\n" else (.choices[0].delta.content? // empty) end'
See the API reference for authentication, the complete schema, polling, and error handling.













