Gemini 4 Carbon leads on hard reasoning and security, but Argon stays the cheaper default

Asked:
“do a good research on google carbon 4 and it should be data driven inculding benchmarks and the graphs and comparisons also”

We read “Google Carbon 4” as Google DeepMind’s Gemini 4 Carbon, the frontier reasoning model announced October 28, 2026 and generally available November 10, 2026. This report draws on 85 evidence rows from 24 pages across 6 hosts: 14 official paired benchmark comparisons from Google, 70 independent benchmark observations (LMArena, Artificial Analysis, Vals AI), and one consolidated official specification row. First-party Google claims and independent evaluations are kept visually and analytically separate, and scores from unlike benchmarks are never averaged or compared directly.

Official paired benchmarks: Carbon vs named comparators

All results below are Google’s own testing (first-party). Panels separate higher-is-better scores from attack-success rates, where lower is better. Dumbbells show comparator → Carbon; exact values labelled. MKA rows carry ± intervals that overlap in places — treat near-ties as statistical ties.

Gemini 4 CarbonComparator (named per row)Carbon aheadComparator ahead
Google’s own MKA paper puts Carbon at 96.36 ±0.91 on Regression-WSR vs GPT-6 Astra’s 95.17 ±1.07 — a lead, but the intervals nearly touch; on Climb-WSR and Overall WSR Claude Fable 5.2 actually edges Carbon. storage.googleapis.com
Security is the clearest win: Gray Swan indirect prompt-injection attack success is 0.0% for Carbon vs 0.7% for Argon, and simulated-attack success falls from 3.1% to 0.8% (lower is better). blog.google
Agentic honesty jumps from 70% (Argon) to 92% (Carbon) on Google’s 375-task benchmark — the largest paired gap in the official set. blog.google
Google also reports Carbon at 100% on Project Fortify (no Argon pair published), and separately states Argon remains the better default for most software-engineering and knowledge work. blog.google

Independent benchmarks: strong, but thinking level matters

Independent evaluators score Carbon by thinking level (Low / Medium / High). Panels are small multiples per evaluator — scales differ and are not comparable across panels. Artificial Analysis puts Carbon High at 60 on its Intelligence Index but Carbon Low at 54; on some tasks (CritPt v2 hard, GDP.pdf) Medium beats High.

Artificial Analysis — task accuracy, % (High vs Medium)
Vals AI — accuracy, % (single Carbon configuration)
Artificial Analysis Intelligence Index: Carbon High 60 (rank 1 among Google models, rank 4 on the open leaderboard), Medium 58, Low 54 — the configuration you buy changes the tier you get. artificialanalysis.ai
Vals AI puts Carbon 2nd on its overall Vals Index (70.4%), with standout Vibe Code Bench 89.1% and Code Migration 80.1%, but weak Tax Agent Bench 48.9% and HLAB 17.5%. vals.ai
Medium beats High on CritPt v2 hard physics (69% vs 63%) and GDP.pdf document finance (30% vs 28%) — more thinking is not uniformly better. artificialanalysis.ai

LMArena: rank 1 in most text categories, with uncertainty attached

Crowd-preference Elo for Gemini 4 Carbon (High), with the published ± interval and vote count. Fresh leaderboards move, and smaller vote counts widen intervals — WebDev Arena (rank 4) and Creative Writing (rank 3) are the exceptions to its rank-1 run.

API economics: a one-million-token window, and prices double on December 1

Context window (shared by input, thinking, output)
1,048,576 tokens
Max output
up to remaining window
Access
Paid tier only — Gemini API / AI Studio; not on free tier
Input, per 1M tokens
$2 → $4
Output + thinking, per 1M tokens
$10 → $20
Cached input, per 1M tokens
$0.10 → $0.20

Introductory pricing runs through November 30, 2026; Google states prices rise December 1. ai.google.dev Artificial Analysis measures effective spend: $7.76 per Intelligence Index task for High and $6.59 for Medium, with High taking 28.3 minutes per task. artificialanalysis.ai

Where Carbon fits — and where Argon still wins

Strengths

Hard, long-horizon reasoning; agentic coding (LMArena coding rank 1, Vibe Code 89.1%); cybersecurity defense (0.0% prompt-injection ASR); agentic honesty (92%); 1M-token multimodal knowledge work.

Weaknesses

Slow and expensive at High thinking (100.5s to first token, $7.76/task); Low configuration drops to Index 54; weak spots on tax (48.9%) and HLAB (17.5%); Claude Fable 5.2 edges it on two of three MKA splits in Google’s own paper.

Best-fit workloads

High-value difficult agentic coding, cybersecurity defense, scientific reasoning, and complex multimodal analysis where quality justifies extra thinking tokens and cost.

Prefer Argon when

Routine or high-volume software-engineering and knowledge work — Google itself calls Argon the better default; switch to Carbon only where a measured quality gain on your own tasks offsets roughly 2× list pricing (post-December 1) plus heavier thinking-token spend.

Bottom line

Carbon is the model to reach for when the problem is genuinely hard and the answer is worth paying for. Benchmark it against Argon on your own task distribution before broad deployment, lock in introductory pricing for evaluation before November 30, and keep Argon for everything routine.

Evidence table — all 85 rows

Every row behind this report: official Google comparisons, independent observations, and the specification row. Blank cells are values the source did not publish. Near-duplicate observations (same metric reported on two pages) are kept here even where the narrative uses one.

TypeSourceCategoryMetricConfigCarbonComparatorComp. scoreRanknBetterLink

Method: 85 evidence rows gathered from 24 URLs across 6 hosts (blog.google, ai.google.dev, storage.googleapis.com, lmarena.ai, artificialanalysis.ai, vals.ai), covering Gemini 4 Carbon’s October 28 announcement through post-GA leaderboard snapshots (LMArena published 2026-10-27; Google posts Oct 28–Nov 11, 2026). 14 rows are Google’s own paired comparisons, 70 are independent benchmark observations, 1 is the official API specification; the largest single URL backs 15 rows (17.6%). Scores measure what each benchmark defines — accuracy %, Elo with ± interval, index points, attack-success rate (lower better), or USD — and are never averaged across benchmarks. Missing uncertainty, sample sizes, and dates are left blank. For space, the charts suppress duplicate observations and non-numeric spec strings; all rows remain in the table.

This report was generated automatically by Keenable SELECT at a user's request, from publicly available web sources linked herein. Keenable does not review, verify, or endorse its contents and makes no representation as to accuracy, completeness, or timeliness; AI-based extraction may contain errors. Nothing in this report is investment, legal, financial, or other professional advice. All trademarks and referenced content remain the property of their respective owners; no affiliation or endorsement is implied. To report an error, rights concern, or request removal: legal@keenable.ai.

Keenable SELECTAsk your own question
Made with Keenable SELECT