Benchgen
Models/google/

Gemini 4 Carbon

DraftPublic

Model Details

Gemini 4 Carbon

Organization License Modality Released Context

Quick answer: Gemini 4 Carbon is Google's follow-up to Gemini 4 Argon, released October 28, 2026. Evaluated independently by Vals AI, it ranks #2 of 32 models on the Vals Index with 70.45% — behind Claude Sonnet 5.5 (70.85%) and ahead of Claude Opus 5.5 (70.18%) — and ranks #1 on both the Vals Legal Research Bench (55.77%) and the Vals Finance Index (67.05%). It has a 1M-token context window (1M max output tokens) and is priced at $4.00/$20.00 per million input/output tokens.

At a Glance

Where it leads: #1 of 89 on Vals Legal Research Bench (55.77%), #1 of 35 on Vals Finance Index (67.05%), and #2 of 32 overall on the Vals Index (70.45%) — all independently measured by Vals AI, not self-reported.

Where it lags: Trails Claude Sonnet 5.5 on the overall Vals Index (70.45% vs. 70.85%) and scores weakest on HLAB-Vals-1.1 (17.50%, tied for 6th of 30) among the benchmarks Vals AI evaluated.

Availability: Available via Google's standard API surface (per Vals AI's evaluation); priced at $4.00 per million input tokens and $20.00 per million output tokens.

What Gemini 4 Carbon Is

Gemini 4 Carbon is Google's frontier model released October 28, 2026, just under a month after Gemini 4 Argon. Unlike Argon's self-reported benchmark table, Carbon's headline numbers here come from Vals AI's independent third-party evaluation suite (Vals Index v2.2), which runs the same proprietary and academic benchmarks across all evaluated models under a consistent methodology.

On the Vals Index — a GDP-weighted composite across finance, coding, legal, and tax tasks — Carbon scores 70.45%, placing #2 of 32 models evaluated, just behind Claude Sonnet 5.5 (70.85%) and just ahead of Claude Opus 5.5 (70.18%) and its own predecessor Gemini 4 Argon (69.24%, #4). Carbon's strongest showings are in research- and finance-oriented work: #1 of 89 on Legal Research Bench (55.77%) and #1 of 35 on the Vals Finance Index (67.05%), per Vals AI's October 28, 2026 evaluation note. It also posts a solid coding result, #4 of 57 on Terminal-Bench 4.0 (63.13%) and 89.06% on Vibe Code Bench (#8 of 121).

Carbon's weakest independently-measured result is on HLAB-Vals-1.1 (Harvey's Legal Agent Benchmark for long-horizon legal work product creation), where it scores 17.50%, tied for 6th of 30 models.

Specifications

FieldValue
OrganizationGoogle
LicenseProprietary (API only)
ModalityMultimodal (text, image, video, file input)
Release date2026-10-28
Context window1,000,000 tokens
Max output tokens1,000,000
Pricing$4.00 / $20.00 per million input/output tokens

Public Benchmark Scores

BenchmarkScoreSourceDate
Vals Index70.45% (#2 of 32)Vals AI2026-10
Legal Research Bench55.77% (#1 of 89)Vals AI2026-10
Terminal-Bench 4.063.13% (#4 of 57)Vals AI2026-10

Scores above are independent third-party measurements by Vals AI, not self-reported by Google. Vals AI's full evaluation also reports Carbon at #1 of 35 on Vals Finance Index (67.05%), #3 of 33 on Vals Legal Index (75.93%), #2 of 37 on EMB v1.1 (76.47%), and 89.06% on Vibe Code Bench (#8 of 121) — those component benchmarks do not yet have standalone Benchgen pages and are not added here to avoid inventing new entities without an existing schema fit.

Frequently Asked Questions

What is Gemini 4 Carbon? Gemini 4 Carbon is Google's frontier model released October 28, 2026, the successor to Gemini 4 Argon released about a month earlier.
How does Gemini 4 Carbon rank on the Vals Index? Gemini 4 Carbon ranks #2 of 32 models on the Vals Index with 70.45%, independently evaluated by Vals AI — behind Claude Sonnet 5.5 (70.85%) and ahead of Claude Opus 5.5 (70.18%).
What is Gemini 4 Carbon's context window? Gemini 4 Carbon has a 1,000,000 token context window and supports up to 1,000,000 output tokens, per Vals AI's evaluation configuration.
How much does Gemini 4 Carbon cost? Gemini 4 Carbon is priced at $4.00 per million input tokens and $20.00 per million output tokens, per Vals AI's published pricing.