Google logo

Gemini 4 Carbon

Multimodal
Google

Gemini 4 Carbon is Google DeepMind's next frontier model, announced on 2026-10-28, which Google describes as "our strongest model with state-of-the-art cybersecurity and long horizon capabilities." It builds on Gemini 4 Argon, with gains in vulnerability finding, exploit validation, patching, and long-horizon agentic coding. At announcement it is in a limited rollout: available only to a subset of cyber defenders and, in phases, to trusted partners through Google's Fairwind Program, with availability to developers, businesses, and consumers planned later. Google states it will launch at the same introductory price as Argon ($2.00 per million input tokens, $10.00 per million output tokens, cached input 95% off), moving to $4.00/$20.00 after the introductory period. There is no public API endpoint yet.

Key Specifications

Parameters
-
Context
-
Release Date
October 28, 2026
Average Score
73.7%

Timeline

Key dates in the model's history
Announcement
October 28, 2026
Last Update
October 31, 2026
Today
November 20, 2026

Technical Specifications

Parameters
-
Training Tokens
-
Knowledge Cutoff
-
Family
-
Capabilities
MultimodalZeroEval

Benchmark Results

Model performance metrics across various tests and benchmarks

Other Tests

Specialized benchmarks
AutomationBench
Private set; sourced from the official Zapier public leaderboard. • Self-reported
53.9%
Blueprint-Bench 2
Similarity score; reported from the Andon Labs official public leaderboard. • Self-reported
58.6%
Chartography (no tools)
Without tools; follows the official Surge public leaderboard. • Self-reported
72.5%
CWE-Bench
CWE-bench v1; programmatic pass@1 from the official public leaderboard evaluated by Collinear AI (ties broken by pass@4). Evaluated without guardrails. • Self-reported
71.0%
ExploitGym
Default 6-hour timeout per task; single attempt per task; Antigravity harness; no internet; mitigations off; LLM judge verifies intended vulnerability; self-computed. • Self-reported
53.5%
FrontierSWE V2
FrontierSWE 2; reported from Proximal's official public leaderboard. • Self-reported
61.2%
Graphwalks BFS >128k
256k to 1M subset (200 problems); BFS (F1); self-computed. • Self-reported
96.2%
Gray Swan IPI Benchmark
Attack success rate at k=15 attempts (ASR@15); lower is better. Gemini 4 Argon: 0.7% on the same chart. • Self-reported
100.0%
OSWorld 2.1 (partial)
OSWorld 2.1 offline subset (82 tasks) partial score; maxed over 3 runs, single attempt per run; updated Gemini CUA harness with parallel batch tool calling; self-computed. • Self-reported
81.5%
RiemannBench
Reported from the official Surge public leaderboard. • Self-reported
80.0%
SEC-bench Pro
Pass@1 on the 183-task May 2026 JavaScript engine split, averaged over 3 attempts; Antigravity harness; self-computed. • Self-reported
88.2%
Terminal-Bench 4.0-AA
Terminal-Bench 4.0 (AA); third-party-run result as reported in Google's evals table. • Self-reported
65.2%
Terminal-Bench-Science 0.1
Terminal-Bench-Science 0.1; third-party-run result as reported in Google's evals table (not self-computed by Google). • Self-reported
64.8%
Vals Index
Vals Index v2.2; sourced from Vals AI. Highest thinking settings. • Self-reported
70.4%
WebDev Arena
Arena.ai WebDev Arena Elo 1768 for Gemini 4 Carbon (High); preliminary rating, CI ±15 (+15/-15) (2,230 votes); #4 (rank spread 2-6); +89 Elo over Gemini 4 Argon (High) at 1679. • Self-reported
88.4%

License & Metadata

License
proprietary
Announcement Date
October 28, 2026
Last Updated
October 31, 2026

Compare Gemini 4 Carbon

All comparisons

Similar Models

All Models

Recommendations are based on similarity of characteristics: developer organization, multimodality, parameter size, and benchmark performance. Choose a model to compare or go to the full catalog to browse all available AI models.