Gemini 4 Carbon Text Generation

Gemini 4 Carbon (gemini-4-carbon) is Google’s strongest reasoning model. Announced on October 28, 2026 and opened to all paying developers and enterprises on November 10, it is now available on APIYI. It is not built for casual chat: it targets long-horizon agentic coding, dense multimodal knowledge work, scientific reasoning and cybersecurity defense. The same model can run build-and-verify loops across a large codebase, or synthesize grounded answers across PDFs, spreadsheets and slide decks. APIYI exposes both the Gemini-native and OpenAI-compatible endpoints in the default and svip groups. Before launch we ran a low-concurrency test covering both endpoints, streaming, function calling, structured output, image input, thinking levels and a needle-in-a-haystack test at roughly 199K tokens. The results and the current limitations are documented on this page.

Pricing: $2 input / $10 output per 1M tokens, Google’s limited-time rate through November 30, 2026. From December 1 the standard $4 / $20 applies and APIYI will follow. Cached input is $0.10 per 1M tokens.

Two things to keep in mind at launch:

Google retains the ability to switch the answering model per request; this did not reproduce after our resource adjustment. If modelVersion is not gemini-4-carbon, log it and retry. It does not mean your request violated any policy; see “Provider-side model switching” below.
maxOutputTokens is now enforced, but it includes thinking tokens. Set it too low (300 did it in our test) and the model can finish thinking with MAX_TOKENS and no answer; leave headroom for hard tasks.

Highlights

Ranked #1 on intelligence

Google reports a score of 59 on the Artificial Analysis Intelligence Index v5, ranked first. It also ranks first on CursorBench 4.0, Perplexity WANDR and Zapier AutomationBench, with Google claiming more than 2x lower cost per task than the next-best external model.

State of the art in cybersecurity

88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for validating exploitability, and 71.0% on CWE-bench for root-cause patching, all Google-reported and evaluated without guardrails.

Official 1M unified window

Input, thinking and output share 1,048,576 tokens; a single streaming segment can emit up to 65,536 tokens, with long-decode continuation beyond that. A 199K-token document passed our needle test; validate 500K–1M with a small sample before relying on it.

Two endpoints, full toolchain

Function calling with round-trips, JSON Schema, image input and streaming all passed on both endpoints. Google also supports code execution, URL context, Computer Use, and Office / PDF file input.

Model information

Parameter	Value
Model name	`gemini-4-carbon`
Input modalities	Text, image, video, audio (official; text and image covered in our test)
Output modalities	Text
Context window	Official 1,048,576 tokens shared by input, thinking and output; 199K tokens verified
Max output	Official 1,048,576 tokens with long decode; 65,536 per streaming segment, 32,768 non-streaming
Thinking levels	`medium` by default; `low` / `medium` / `high` supported; `thinkingBudget` and `minimal` are not supported
Knowledge cutoff	March 2026
Groups	`default`, `svip`
Endpoints	`POST /v1beta/models/gemini-4-carbon:generateContent` (native), `POST /v1/chat/completions` (OpenAI-compatible)
Streaming	✅ Both endpoints
Caching	Implicit caching only (official, 4,096-token minimum); cached input $0.10 per 1M tokens

Official benchmarks

All figures below are published by Google and were not independently reproduced by APIYI:

Benchmark	What it measures	Gemini 4 Carbon
Artificial Analysis Intelligence Index v5	General intelligence	59 (#1)
SEC-bench Pro	Finding real vulnerabilities	88.2%
ExploitGym	Validating exploitability	54.8%
CWE-bench	Root-cause patching	71.0%
WebDev Arena	Web development	#4, 1768 (+89 Elo)
CursorBench 4.0	Real-world coding tasks	#1
Perplexity WANDR	Wide-and-deep research	#1
Zapier AutomationBench	End-to-end business workflows	#1, >2x lower cost per task
Artificial Analysis MKA	Multimodal knowledge agents	#2, >6x lower cost per task

Third-party Artificial Analysis separately measured an AA-Omniscience hallucination rate of 11% for Carbon (Argon: about 4%) and calls it very verbose (about 200M output tokens for a full index run versus a 64M median). For factual work, ask for sources and verify.

Thinking and parameters

Setting	Behavior
No `thinkingConfig`	Defaults to `medium`
`thinkingConfig.thinkingLevel: low`	Least thinking, fastest first token; simple tasks and writing
`thinkingConfig.thinkingLevel: medium`	Google’s recommendation for most code and agent work
`thinkingConfig.thinkingLevel: high`	Most thinking, slowest first token; hard math, science, complex code
`thinkingConfig.thinkingBudget`	Not supported by Google and should return 400; a value of 0 was sometimes accepted in our test, do not rely on it
`thinkingConfig.thinkingLevel: minimal`	400, `Thinking level MINIMAL is not supported`
`temperature` / `topP` / `topK`	Fixed at 1.0 / 0.95 / 64 by Google; custom values are ignored
`maxOutputTokens`	Official cap for thinking + output; enforced now, but too small a value can leave no answer
`stopSequences`	✅ Honored (cut at `END` in our test)

maxOutputTokens includes thinking tokens, so Google recommends leaving enough headroom. In our test, 50 / 100 / 300 all ran out during thinking and returned finishReason=MAX_TOKENS with little or no text. Set it to a few thousand for a complete answer; the parameter is now passed through by APIYI.

Context and output: official 1M, 199K verified

The official model card lists 1,048,576 tokens shared by input, thinking and output, with up to 65,536 tokens per streaming segment and long-decode continuation beyond that. After the resource adjustment we re-tested:

Input size	Reported `promptTokenCount`	Needle at the end found
616,090 characters (~199K tokens)	199,277	✅

The earlier ~128K cap and silent truncation did not reproduce. We did not test the 500K / 1M range in this round; validate with a small sample before sending very long documents.

Quickstart

Gemini-native endpoint

curl https://api.apiyi.com/v1beta/models/gemini-4-carbon:generateContent \
  -H "Authorization: Bearer $APIYI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "contents": [{"role": "user", "parts": [{"text": "Prove that the square root of 2 is irrational."}]}],
    "generationConfig": {
      "thinkingConfig": {"thinkingLevel": "high"}
    }
  }'

OpenAI-compatible endpoint

curl https://api.apiyi.com/v1/chat/completions \
  -H "Authorization: Bearer $APIYI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemini-4-carbon",
    "messages": [{"role": "user", "content": "Explain tidal locking."}],
    "max_tokens": 1024
  }'

Note: max_tokens is now enforced, but thinking tokens count toward it. Too small a value produces a MAX_TOKENS response with no answer; leave headroom for hard tasks.

Pricing and caching

Item	Price per 1M tokens
Input	$2 ($4 from December 1, 2026)
Output (including thinking)	$10 ($20 from December 1, 2026)
Cached input	$0.10 (95% off input)

The limited-time rate runs through November 30, 2026; from December 1, Google’s standard price applies and APIYI will follow.
Implicit caching only, 4,096-token minimum; explicit cachedContents is not supported.
Thinking tokens are billed as output, so high costs noticeably more than low.

Provider-side model switching: what it looks like and what to do

What happens: when you call gemini-4-carbon, the native response’s modelVersion should almost always be gemini-4-carbon. We previously saw some responses report 3.1-flash-lite-05-2026, with much lower latency and very few thinking tokens; in streaming, the same responseId could carry both versions across chunks. After the resource adjustment, this did not reproduce. How to read it: this is Google switching the answering model per request. Google’s usage policies allow changing the answering model per request, and Google’s developer forum has a closely matching report of gemini-3.1-pro-preview being served as 3.1-flash-lite-preview-03-2026; the snapshot we saw on APIYI was 05-2026. It does not mean your request violated anything and does not mean your account was penalized. The exact trigger cannot be confirmed from outside; it may be routing or risk-control policy, or a version or metering metadata anomaly. It is stable now, but Google retains the ability. Impact:

Simple tasks usually look the same; hard tasks, long-horizon agents and deep-reasoning work can lose quality when served by Flash-Lite.
If it happens again, latency will look unusually good and thinking tokens will be sparse; do not treat that as Carbon-native performance.

Recommendations:

For critical work, log modelVersion from native responses (the OpenAI-compatible endpoint does not return it) and decide whether to retry when it reports 3.1-flash-lite.
For consistency-sensitive work, add verification: ask for sources, run unit tests, or cross-check with another model.
Use thinkingLevel: high and explicit step constraints to make the task more resilient.
Do not put “always served by Carbon” into an SLA.

Choosing a model

Scenario	Recommendation
Hard problems, long-horizon coding, scientific reasoning, security auditing	Carbon with `thinkingLevel: high`, with enough output budget
Everyday coding, knowledge work, interactive tasks	Argon is faster and cheaper; Carbon’s extra thinking may not pay off
High concurrency, low latency, cost-sensitive	Flash series or Flash-Lite; Carbon is not a fit
Very long documents (>200K tokens)	Only 199K was verified this round; chunk them or validate with a small sample first

Model pricing is aligned with the official site and may change; the table above is for reference only. The Model Pricing tab in the top navigation is authoritative: Model Pricing.

Basics

Basic API

Image API (Official)

Video API (Official)

Multimodal Understanding API

Text Generation

Audio API (Official)

Retrieval & Safety

Highlights

Ranked #1 on intelligence

State of the art in cybersecurity

Official 1M unified window

Two endpoints, full toolchain

Model information

Official benchmarks

Thinking and parameters

Context and output: official 1M, 199K verified

Quickstart

Gemini-native endpoint

OpenAI-compatible endpoint

Pricing and caching

Provider-side model switching: what it looks like and what to do

Choosing a model

​Highlights

Ranked #1 on intelligence

State of the art in cybersecurity

Official 1M unified window

Two endpoints, full toolchain

​Model information

​Official benchmarks

​Thinking and parameters

​Context and output: official 1M, 199K verified

​Quickstart

​Gemini-native endpoint

​OpenAI-compatible endpoint

​Pricing and caching

​Provider-side model switching: what it looks like and what to do

​Choosing a model

​Related docs

Highlights

Model information

Official benchmarks

Thinking and parameters

Context and output: official 1M, 199K verified

Quickstart

Gemini-native endpoint

OpenAI-compatible endpoint

Pricing and caching

Provider-side model switching: what it looks like and what to do

Choosing a model

Related docs