Google

Gemini 4 Carbon

Google's strongest Gemini 4 model, built on Argon for complex software engineering, long-running agents and hard reasoning.

built on gemini-4-argontext inputtext output

Run this model

google/gemini-4-carbon

Use an endpoint supported by this model.

Get a key for this setup →
Thinking guide →
const response = await fetch('https://api.minirouter.sh/v1/chat/completions', {  method: 'POST',  headers: {    Authorization: `Bearer ${process.env.MINIROUTER_KEY}`,    'Content-Type': 'application/json',  },  body: JSON.stringify({    model: 'google/gemini-4-carbon',    messages: [{ role: 'user', content: 'Why is the sky blue?' }],    max_tokens: 1024,  }),}) const data = await response.json()console.log(data.choices[0].message.content)
Read docs →

Set max_tokens to at least 32,768: thinking shares that budget. Start at medium effort; a streamed reply must begin within 95 seconds.

Introductory provider pricing until November 30, 2026Current prices include MiniRouter's markup. Google lists $2 input, $10 output and $0.10 cached input per million tokens through November 30, 2026, then $4 input and $20 output per million tokens from December 1, 2026. Reasoning counts as output; final charges follow provider-reported usage. Source ↗

Model overview

Input / 1M
$2.1
All-in MiniRouter rate
Output / 1M
$10.5
All-in MiniRouter rate
Context
1M
1,000,000 tokens
Max output
1M
tokens
Effort
low → high
medium by default, always thinking
Intelligence Index
60
Artificial Analysis, high effort
Announced
28 Oct 2026
publisher release

Compare text-model alternatives or estimate token costs for your workload.

Publisher-reported, plotted by MiniRouter

Ahead on long context, agents and research.

Carbon leads seven of these nine rows from Google's launch table. GPT-6 Astra leads FrontierSWE 2 and Claude Opus 5.5 leads OSWorld-2.1. MiniRouter plotted the table and did not run the evaluations.

Long-context reasoning

Keeps its place across a million tokens.

200 graph-search problems hidden in prompts of 256K to 1M tokens, scored by F1. Google ran all five models itself.

Gemini 4 Carbon
96.2%Highest thinking
Gemini 4 Argon
84.2%Highest thinking
GPT-6 Astra
71.8%Run by Google

GraphWalks · 256K to 1M tokens

0 to 100%

  • Gemini 4 Carbon96.2%
  • Gemini 4 Argon84.2%
  • GPT-6 Astra71.8%
  • Claude Opus 5.566.8%
  • Claude Fable 5.165.0%
Source: Google model card ↗ · Evaluation methods ↗. Results as of October 2026.
View chart data
GraphWalks · 256K to 1M tokens scores
ModelF1
Gemini 4 Carbon96.2%
Gemini 4 Argon84.2%
GPT-6 Astra71.8%
Claude Opus 5.566.8%
Claude Fable 5.165.0%

Eight more evaluations

Agents, coding, research and professional work

  • Gemini 4 Carbon
  • Comparison models

Long-running agents

Agents' Last Exam v1.1

  • Gemini 4 Carbon77.4%
  • Gemini 4 Argon69.0%
  • GPT-6 Astra73.1%
  • Claude Opus 5.571.6%
  • Claude Fable 5.1—Not reported

0 to 100%

Agentic coding

FrontierSWE 2

  • Gemini 4 Carbon61.2%
  • Gemini 4 Argon55.0%
  • GPT-6 Astra65.5%
  • Claude Opus 5.562.3%
  • Claude Fable 5.156.3%

0 to 100%

Scientific research

Terminal-Bench-Science 0.1

  • Gemini 4 Carbon64.8%
  • Gemini 4 Argon52.9%
  • GPT-6 Astra63.3%
  • Claude Opus 5.559.0%
  • Claude Fable 5.143.3%

0 to 100%

Business workflows

AutomationBench

  • Gemini 4 Carbon53.9%
  • Gemini 4 Argon51.3%
  • GPT-6 Astra41.4%
  • Claude Opus 5.542.5%
  • Claude Fable 5.131.4%

0 to 100%

Computer use

OSWorld-2.1 · offline subset

  • Gemini 4 Carbon81.5%
  • Gemini 4 Argon74.7%
  • GPT-6 Astra80.9%
  • Claude Opus 5.587.2%
  • Claude Fable 5.176.8%

0 to 100%

Professional work

Vals Index v2.2

  • Gemini 4 Carbon70.4%
  • Gemini 4 Argon69.2%
  • GPT-6 Astra65.0%
  • Claude Opus 5.570.2%
  • Claude Fable 5.1—Not reported

0 to 100%

Mathematics

RiemannBench

  • Gemini 4 Carbon80.0%
  • Gemini 4 Argon76.0%
  • GPT-6 Astra72.0%
  • Claude Opus 5.569.6%
  • Claude Fable 5.165.6%

0 to 100%

Chart reading

Chartography · no tools

  • Gemini 4 Carbon72.5%
  • Gemini 4 Argon71.6%
  • GPT-6 Astra71.0%
  • Claude Opus 5.566.3%
  • Claude Fable 5.146.2%

0 to 100%

Gemini models at their highest thinking setting. Other models' scores are their providers' own or from public leaderboards, at maximum reasoning where available. Source: Google model card ↗ · Evaluation methods ↗. Results as of October 2026.
View chart data
Values from Google's Gemini 4 Carbon launch table
BenchmarkGemini 4 CarbonGemini 4 ArgonGPT-6 AstraClaude Opus 5.5Claude Fable 5.1
GraphWalks · 256K to 1M tokens96.2%84.2%71.8%66.8%65.0%
Agents' Last Exam v1.177.4%69.0%73.1%71.6%Not reported
FrontierSWE 261.2%55.0%65.5%62.3%56.3%
Terminal-Bench-Science 0.164.8%52.9%63.3%59.0%43.3%
AutomationBench53.9%51.3%41.4%42.5%31.4%
OSWorld-2.1 · offline subset81.5%74.7%80.9%87.2%76.8%
Vals Index v2.270.4%69.2%65.0%70.2%Not reported
RiemannBench80.0%76.0%72.0%69.6%65.6%
Chartography · no tools72.5%71.6%71.0%66.3%46.2%
Choose an effort

High thinks far longer before it answers.

Medium is Google's default and scores within two points of high. At high effort a request can time out before the first token arrives, so start with medium.

Intelligence Index v5

Low
54.2
Medium
58.3
High
60.0

Wait before the answer starts

Seconds to first answer token

0 to 120 seconds

  • Low3.8
  • Medium4.9
  • High100.8
Artificial Analysis ↗, measured on Google's API and read November 10, 2026. These are not MiniRouter route measurements.
View chart data
Gemini 4 Carbon by effort, measured by Artificial Analysis
EffortIntelligence IndexFirst answer token
Low54.23.8 s
Medium58.34.9 s
High60.0100.8 s
Publisher pricing

Argon's price, and it doubles on December 1.

Google lists the same per-token rates for Carbon and Argon. Carbon spends more thinking tokens per turn, so the same task costs more. The price panel above shows what you pay on MiniRouter.

Google list, per 1M tokens, through November 30

Input
$2.00$4.00 from December 1
Output
$10.00$20.00 from December 1
Cached input
$0.10$0.20 from December 1
List prices from Google's pricing page ↗, before MiniRouter's fee. Output includes thinking tokens.
View chart data
Google list prices for Gemini 4 Carbon per million tokens
RateThrough November 30, 2026From December 1, 2026
Input$2.00$4.00
Output$10.00$20.00
Cached input$0.10$0.20
Before a long run

Four behaviours to plan for.

Google describes the first three in its Carbon guide. Argon requests need no other change than the model ID.
  1. [ 01 ]

    The output cap counts thinking

    max_tokens covers thinking plus the visible answer. Google recommends at least 32,768.

  2. [ 02 ]

    Thinking cannot be turned off

    Choose low, medium or high. Google's default is medium.

  3. [ 03 ]

    Sampling parameters are ignored

    Google ignores temperature, top_p and top_k on Carbon.

  4. [ 04 ]

    A streamed reply must start within 95 seconds

    MiniRouter ends a streaming request that has produced nothing by then. High effort can take longer; medium and low answer in seconds.

Route

One route, standard clients.

Authenticate with a MiniRouter bearer key. Caller-provided upstream credentials are not used on this route.
Base URLhttps://api.minirouter.sh/v1
Model IDgoogle/gemini-4-carbon
Endpoints/v1/chat/completions/v1/messages/v1/responses
Parametersmax_tokens, temperature, stop, tools, tool_choice, reasoning, include_reasoning
Structured outputNot enforced; a requested format is dropped and the answer is plain text.
ModalitiesInput: text · Output: text

Configured providers. Priority follows the live catalog order.

ProviderPriorityInput / 1MOutput / 1MParametersAccess
vercel
1$2.1$10.5max_tokens, temperature, stop, tools, tool_choice, reasoning, include_reasoningMiniRouter balance

Every model, one key.