Gemini 4 Carbon

Google·2026-10-28

LLMsflagshipcloud

Overview

Second Gemini 4 model, built on Gemini 4 Argon and positioned by Google as its strongest model for cybersecurity and long-running engineering tasks. Restricted access at launch: select cyber-defense organizations get it for pre-release testing while Google refines guardrails before a broad release; no public API, playground or weights on announcement day. 1M-token context and up to 1M output tokens. Introductory API pricing of $2 / $10 per MTok with 95% off cached input; Vals AI and Artificial Analysis list $4 / $20 on their own runs. Vendor-reported cyber figures (SEC-bench Pro 88.2, ExploitGym 54.8, CWE-bench 71.0) are not tracked benchmarks and, per launch coverage, were run without guardrails. Google's launch post was not reachable from this environment, so no vendor figure is carried over here (checked 2026-10-31).

Capabilities and innovations

Cyber defense (vulnerability discovery and patching, secure code generation)Long-horizon agentic software engineering1M-token context, up to 1M output tokensMultimodal input (text, image, video, files)Built on Gemini 4 ArgonRestricted pre-release access for cyber defenders before general availability

Benchmarks

Benchmark	Score	Source
Terminal-Bench 4.0	63.13%	independent Terminal-Bench 4.0 63.13% (±2.20), measured by Vals AI, reasoning effort high, default temperature, rank 4/59 (accessed 2026-10-31). Artificial Analysis measures 65.15 on its own run; launch coverage quotes 65.2 in Google's comparison. Same source and setting as the Argon entry (57.58). 4.0 is its own generation and not comparable with the 2.x or 3.0 fields.

GPQA Diamond: independent evaluation pending — Artificial Analysis shows no GPQA Diamond value and Vals AI's model page lists none (checked 2026-10-31).
Humanity’s Last Exam (no tools): independent evaluation pending — Artificial Analysis has no HLE value for Carbon (checked 2026-10-31). Its 69.1 is on HLE Diamond, a different subset, and is not written into the HLE field.
SWE-bench Verified: not reported by the vendor
SWE-bench Pro: not reported by the vendor — Not in any launch coverage found; Google led Argon with DeepSWE v1.1 instead.
MMLU: not reported by the vendor
MMLU-Pro: independent evaluation pending — No Vals AI or Artificial Analysis value yet (checked 2026-10-31).
Terminal-Bench 2.x: not reported by the vendor — Vals AI and Artificial Analysis measure 4.0 only, recorded in terminal_bench_4.
Terminal-Bench 3.0: not reported by the vendor
DeepSWE v1.1: independent evaluation pending — No v1.1 figure for Carbon found in launch coverage; the 77.9 circulating in coverage belongs to Argon. Google's launch post and Datacurve's board were not reachable from this environment (checked 2026-10-31).
MMMU-Pro: independent evaluation pending — Artificial Analysis lists no MMMU-Pro value yet (checked 2026-10-31).

API pricing

$2 per million input tokens, $10 per million output tokens (USD, provider list price, checked 2026-10-31)

Reliability

omniscience_index: 53.5 ()independent

More from Google

Data curated by AI Model Timeline. See the methodology for admission criteria, benchmark eras and source priority.