LLM Benchmark Snapshot, November 3 2026: Gemini 4 Carbon Takes the Text Arena

Our benchmarks page takes a snapshot of the public LLM leaderboards every time the daily runs, and each snapshot keeps its own permanent URL. This is the photo for the week of November 3, 2026, compared against the last weekly post (September 22) and the October 25 snapshot.

The headline: Gemini 4 Carbon takes the text arena

On the LMArena text leaderboard, Gemini 4 Carbon (High) has held the number one spot in every snapshot since October 29, at roughly 1540 Elo. It replaced its sibling Gemini 4 Argon (High), which had led through most of October and now sits second at about 1529. Claude Opus 5.5 (High) is third at about 1516.

For context, the September 22 snapshot had the Claude Fable 5.1 and Claude Opus 5 variants at the top, all around 1505. In six weeks the top score moved up by about 35 points, and the leader changed family.

BenchLM agrees, narrowly

BenchLM's overall leaderboard also puts Gemini 4 Carbon first (88.4), followed by Claude Opus 5.5 (86.3) and GPT-6 Astra (86.1). On October 25, before Carbon showed up there, Claude Opus 5.5 led with 86.6. The gap between second and third today is a fraction of a point, so treat the order below the leader as a tie.

WebDev is still Claude territory

The LMArena WebDev arena did not change hands: Claude Opus 5.5 (Max) leads at about 1810, with GPT-6 Astra (Max) at about 1773 and Claude Sonnet 5.5 (xHigh) right behind at about 1770. The same top three held on October 25. Back in September, GPT-6 Astra led this board, so Anthropic's lead here dates from the Opus 5.5 release.

Aider polyglot is frozen

The Aider polyglot coding leaderboard has not been updated with new frontier models for weeks: GPT-5 (high) at 88.0, GPT-5 (medium) at 86.7 and o3-pro (high) at 84.9, identical to September. Read it as a historical reference, not a current ranking.

Sources disagree, and that is the point

The same week, three different leaders depending on where you look: Gemini 4 Carbon on general chat preference and on BenchLM's aggregate, Claude Opus 5.5 on web development, and a months-old GPT-5 on a stale coding board. No single number tells you which model to use. Pick the leaderboard that matches your task, and check whether it is still being updated.

One data note: our LiveBench feed has returned no rows since the October 18 snapshot, so LiveBench is missing from this comparison.

Explore the data