Gemini 4

The Gemini 4 family

Gemini 4

HHydrogen CCarbon NeNeon ArArgon

Scroll

Periodic family

Four elements, four models

All four read a 1,048,576-token context.

11.008
HHydrogen1s¹

Gemini 4 Hydrogen

Everyday execution

Announced
Nov 12, 2026
Per 1M tokens, in / out
$1.50 / $7.50
Thinking levels
low → xhigh
Strongest
612.011
CCarbon[He] 2s² 2p²

Gemini 4 Carbon

The hardest problems

Announced
Oct 28, 2026
Per 1M tokens, in / out
$2 / $10 intro
Thinking levels
low → high
1020.180
NeNeon[He] 2s² 2p⁶

Gemini 4 Neon

Depth with efficiency

Announced
Nov 18, 2026
Per 1M tokens, in / out
$3 / $15
Thinking levels
low → xhigh
1839.948
ArArgon[Ne] 3s² 3p⁶

Gemini 4 Argon

Long-horizon reasoning

Announced
Sep 30, 2026
Per 1M tokens, in / out
$2 / $10 intro
Thinking levels
low → high

C · 6 · Mohs 10

Gemini 4 Carbon

Google’s strongest model, built for the hardest problems.

Diamond · n = 2.42 · dispersion 0.044
59Artificial Analysis Intelligence Index v5.0, #1 at launch
Nov 10Generally available. Cyber defenders got it first, from Oct 28

$2 in / $10 out per 1M tokens, introductory until Nov 30, 2026, then $4 / $20

Valence 4 · sp³ 109.5°

One prompt, a coordinated team

/swarm in Google Antigravity, in preview

10overnight loops
906Ktriangles in the rebuilt model
206drawing views pulled from 7 city-council PDFs

Google’s /swarm demo: a reconstruction of Platform 37

Diamond · the hardest natural material

Diamond-hard cyber defense

99%of simulated attack trajectories blocked
0.8%breach rate per attempt, down from 3.1% on Argon
  1. FindSEC-bench Pro88.2%
  2. PatchCWE-bench71.0%
  3. ExploitExploitGym54.8%

Project Fortify: 100% recall, 95% precision. Wiz Scan for Good: 0-days found in under two hours.

C–C bond 154 pm · the element of life

40× more protein homologs

CASP17 blind test set, seven low-homolog families

80.0%RiemannBench, mathematics
64.8%Terminal-Bench-Science 0.1
96.2%GraphWalks, 256K to 1M context

Type IIa · flawless clarity

Knows what it doesn’t know

92%agentic honesty, up from 70% on Argon
98%declined to invent missing experimental data
0%attack success on Gray Swan prompt injection

Trajectory Labs: no new universal jailbreaks across 132 hours and 13,000+ requests.

CH₄ · tetrahedral 109.5°

Carbon plans. Hydrogen executes.

Orchestrator in Google Antigravity switches between the two inside one session.

Task successnear Carbon
Quota used−24%
Time to finish19% faster

Terminal-Bench 4.0: near-Carbon task success with 24% less quota, 19% faster on average.

H · 1 · 1.008 u

Gemini 4 Hydrogen

The daily driver for agentic work.

Balmer series
51 → 54Artificial Analysis Intelligence Index, high → xhigh. Argon: 50
$1.50 / $7.50per 1M tokens, the same price across the full 1M context

The lightest element

2.5× faster

than Gemini 3.8 Flash, at default thinking

57%Terminal-Bench 4.1, up from 18.2% on 3.8 Flash
12–23%fewer tokens in partner tests, at 2× to 2.8× the speed

Agentic tool use

Parallel calls, graceful recovery

Batches file reads and searches, and adapts when a call fails.

>5×fewer fabricated claims than 3.8 Flash
94%honesty rate

Embodied reasoning

From a sentence to a moving arm

Drives a 6-axis arm from a plain-language goal and a camera view.

58Artificial Analysis Robotics 0.1
ERQAahead of Gemini Robotics ER 2 on multi-view and 3D-trajectory tasks

The most abundant element

Everywhere, free tier included

  • Google Search, for everyone
  • Gemini app
  • AI Studio and Gemini API, free tier
  • Google Sheets
  • Android Studio
  • Google Antigravity
58.8%HLE-Diamond
52.3%Vals Tax Agent Bench
71.6%Agents’ Last Exam v1.1

Ne · 10 · 640.2 nm

Gemini 4 Neon

Depth and efficiency, in balance.

Neon emission lines
32%lower cost per task than Carbon, both at high
$3 / $15in / out per 1M tokens

Steerability

Correct it mid-task

Move your pointer across the tube to steer it.

#2on the Agent Arena steerability measure
1336Agent Arena Index score, across 4,200+ sessions

Noble gas · full shell 2s² 2p⁶

Doesn’t cut corners

0.0%cheating on Terminal-Bench 4.0
2.45%shortcuts on Google’s coding cheating eval, the lowest of any Gemini 4 model
99.2%agent honesty pass rate

Lowest fabrication rate recorded so far in Google’s internal Antigravity testing.

Long-running agents

Lights up a team of Hydrogens

Runs parallel Hydrogen subagents, rereads fewer files, compacts its context.

71.3%HLE-Diamond, best in its class
50–80%lower cost per task than top frontier models on Vals Index and Agent Arena

Ar · 18 · 39.948 u

Gemini 4 Argon

Built to sustain deep reasoning across long-horizon work.

Argon emission lines
77.9%DeepSWE v1.1, state of the art at launch. Opus 5.5: 74.2%, GPT-6 Astra: 74.1%
Sep 30announced first of the four

Argon · welding shield gas

Shielded from injected instructions

Leading on Gray Swan’s indirect prompt injection benchmark at launch.

Ar⁺ laser · 488 nm

Finds, validates, patches

68%CWE-bench v1, tied for first
2.7×faster than the existing Rust port of the libgav1 decoder
300+ TiBof memory freed in Google’s own systems

Long decode

Up to 1M output tokens

Up from 64K, sixteen times longer.

91.7%LVBench, long video, state of the art at launch
51.3%AutomationBench (Zapier)
$2 / $10per 1M tokens in / out, introductory. Then $4 / $20

Spectrograph

Benchmarks

The Gemini 4 Carbon evaluation table: 17 benchmarks, 5 models. Each score is an emission line.

    Scroll, or use ↑ ↓

    Methodology: deepmind.google/models/evals-methodology/gemini-4-carbon

    Gemini 4

    Hydrogen · Carbon · Neon · Argon

    Gemini API · Google AI Studio · Google Antigravity · Gemini app