Gemini 4 Carbon Enters Agent Arena on Launch Day as Its Lab-Benchmark Sweep Meets Live Judgment
Gemini 4 Carbon entered three Arena leaderboards on October 28 — Agent Arena, Text Arena, and Code Arena WebDev — the same day Google published its launch post. It is the first time the model faces human-preference evaluation independent of any Google-controlled benchmark run.
The model arrives with an unusually strong static record. On Artificial Analysis Intelligence Index v5.0, Gemini 4 Carbon (High) holds the #1 spot at 59, ahead of Claude Opus 5.5. On Vals Index v2.2, it posts 70.4%. Terminal-Bench 4.0, as evaluated by Artificial Analysis, scores 65.2% — placing it above Grok 4.7 (37.6%) and GPT-5.6 Sol (37.3%), though below Claude Sonnet 5.5 (70.6%) and Claude Opus 5.5 (66.4% at standard effort).
In Code Arena: WebDev, Google’s launch materials report an early ELO of 1,768, placing it at #4 on that leaderboard. Agent Arena and Text Arena ELOs are still accumulating battles.
The Agent Arena Addition Is the One That Counts
Agent Arena differs from Text Arena in a specific way: it scores models on multi-step task completion judged by human evaluators, not a single conversational exchange. For developers choosing a model for autonomous pipelines, it is closer to real signal than any static leaderboard.
Gemini 4 Carbon’s agentic benchmark footprint is mixed. Agents’ Last Exam v1.1 at 77.4% is among the highest recorded. AutomationBench lands at 53.9%, a strong position. FrontierSWE 2 hits 61.2%.
But SWE-sweep — a coding-agent evaluation that measures whether a model can identify and patch vulnerabilities across a broad codebase sweep — sits at 9.6%. That is an anomalously low figure for a model claiming the intelligence index top spot, and it will draw scrutiny in Agent Arena’s coding tasks. CWE-bench (71.0%) and SEC-bench Pro (88.2%) tell a different story on targeted security tasks, which suggests the gap is specific to broad autonomous code-sweep workloads rather than security knowledge in general.
Key Numbers
| Benchmark | Score | Context |
|---|---|---|
| AA Intelligence Index v5.0 | 59 (#1) | Leads Claude Opus 5.5 |
| Vals Index v2.2 | 70.4% | |
| Terminal-Bench 4.0 (AA eval) | 65.2% | Behind Sonnet 5.5 (70.6%) |
| AutomationBench | 53.9% | Strong agentic positioning |
| Agents’ Last Exam v1.1 | 77.4% | Among highest on record |
| FrontierSWE 2 | 61.2% | |
| SWE-sweep | 9.6% | Anomalously low — watch this in Agent Arena |
| Code Arena WebDev ELO | 1,768 (#4) | From Google launch materials; battles ongoing |
Pricing and What Is Missing
Intro pricing is $2/$10 per 1M input/output tokens, rising to $4/$20 after the introductory period. That puts it in the same bracket as Claude Opus 5.5 ($4/$20) at a temporary 50% discount.
What the model lacks is primary benchmark coverage under the new evaluation regime. SWE-bench Verified was deprecated in September 2026 after Claude Opus 5 hit the 97% ceiling, and tau-bench has no Gemini 4 Carbon entry yet. The Stack Futures ticker carries a null composite score for this model until Vals completes a SWE-bench Pro or tau-bench run. Agent Arena ELO, once it stabilises across several hundred battles, will be the most useful independent data point available.
The lab-benchmark versus Arena-ELO divergence has been sharp before — models that lead static indices have landed mid-table on Agent Arena when the task set drifts from their training distribution. Whether Gemini 4 Carbon holds its #1 position against Claude Fable 5.1, GPT-6 Astra, and Claude Opus 5.5 in live battles is the question this Arena entry begins to answer.