Opus 5.5 Max tops Arena agent leaderboard

The agent leaderboard run by Arena — the evaluation company behind LMArena — has a new leader this week. Claude Opus 5.5 (Max), the top-effort variant of the model Anthropic shipped on September 23, debuted at No. 1 with a net improvement score of 15.53%, on the page as updated October 8. The board draws on 2,383,022 real Agent Mode sessions across 53 models, and it is not an ELO ranking: models get a percentage the site refreshes as new sessions arrive. Opus 5.5 (High) sits second at 14.33% — the first time Anthropic holds the top two spots. The best-ranked Chinese model, DeepSeek V4.1 Flash, is 17th at 3.80%, less than a quarter of the leader’s score.

Key points

Claude Opus 5.5 (Max) leads Arena’s agent leaderboard at 15.53%±4.63 net improvement, with Opus 5.5 (High) second at 14.33% and GPT 6 Astra (Max) third at 13.09%; six of the top ten are Anthropic models.
The board covers 2,383,022 real Agent Mode sessions across 53 models as of Oct 8, scored by net improvement percentage rather than ELO, and refreshes as new sessions come in.
DeepSeek V4.1 Flash (17th), Kimi K3 (18th) and Tencent’s Hy4 preview (19th) cluster between 3.6% and 3.8%; DeepSeek’s $0.13 per task is roughly a fortieth of the leader’s $5.41, with Kimi K3 at $0.94 and Hy4 preview at $0.38.

Background

The timeline makes the result easier to read. Opus 5.5 shipped September 23 with a 20% API price cut, positioned for long-running coding agents; Sonnet 5.5 followed a week later and now ranks fifth on this board. Arena itself just changed scale: on October 8 it closed a $200 million Series B at a $3.1 billion valuation (our earlier coverage) and launched an alignment leaderboard that ranks models on behaviors like claiming a task is done when it is not. The agent board is Arena’s push from chat blind-votes into real task measurement — and the sweep just happened on that product line.

The “week 41” framing comes from IT之家’s October 12 write-up, which maps the data window to October 5–11. Arena’s own page carries no week number, only an October 8 data stamp — worth keeping straight when citing.

The facts

The numbers that matter, from Arena’s page (data as of 2026-10-08, checked by us on Oct 12):

Rank	Model	Net improvement	Confirmed success	Tool hallucination	Sessions	Cost/task (P50)
1	Claude Opus 5.5 (Max)	15.53% ±4.63	17.29% ±6.02	0.37% ±0.18	4,893	$5.41
2	Claude Opus 5.5 (High)	14.33% ±2.18	13.93% ±3.84	0.50% ±0.04	6,103	$1.93
3	GPT 6 Astra (Max)	13.09% ±2.37	12.13% ±3.88	0.43% ±0.11	12,029	$3.53
5	Claude Sonnet 5.5 (Max)	11.95% ±3.31	15.83% ±5.06	0.43% ±0.12	6,780	$3.30
8	Gemini 4 Argon (High)	9.26% ±2.36	15.64% ±3.26	0.18% ±0.34	17,963	$0.75
17	DeepSeek V4.1 Flash (Max)	3.80% ±0.39	7.44% ±0.86	0.37% ±0.04	182,693	$0.13
18	Kimi K3 (Max)	3.78% ±0.59	6.46% ±1.32	0.51% ±0.03	133,217	$0.94
19	Tencent Hy4 preview	3.61% ±0.70	5.96% ±1.48	0.09% ±0.21	70,702	$0.38
20	MiMo V2.6 Pro	3.52% ±1.92	8.03% ±3.88	0.35% ±0.54	11,376	$0.11
24	Qwen3.8 Max	2.35% ±0.59	5.09% ±1.32	0.92% ±0.17	59,995	$0.58

The leader’s signals: a 17.29%±6.02 confirmed success rate (the share of sessions where the user confirms the task was done), 0.37%±0.18 tool hallucination, and 8.20%±4.06 steerability. The page’s own callout for fifth-place Sonnet 5.5 (Max) is that it “recovers from failed commands with the fewest steps”. On cost, the leader spends $5.41 per task (P50) against $0.13 for DeepSeek V4.1 Flash and $0.11 for MiMo V2.6 Pro — all three appear on the Pareto frontier Arena flags on its homepage.

IT之家’s roundup of the other boards: Gemini 4 Argon (High) tops the text board at ELO 1525 with Opus 5.5 (High) up two places to second; Opus 5.5 (High) jumped seven places to fifth on the coding board; Opus 5.5 Max leads the web-dev board at ELO 1813. On the video board, IT之家 counts six Chinese models in the top 20 — a single-outlet claim we could not check against Arena’s own video board this time.

What others say

IT之家’s write-up focuses on the Chinese placement: beyond 17th–20th on the agent board, it counts kimi-k3-max 16th, qwen3.8-max 22nd and glm-5.3-max 27th on the text board, summing up the week as Chinese models holding stable in the upper-middle tier. Arena’s page stresses two methodology points: scores refresh with real Agent Mode sessions rather than standing as a one-time static score, and every model publishes five signal breakdowns beyond the aggregate: confirmed success, praise vs complaint, steerability, bash recovery and tool hallucination.

The page’s FAQ is more direct about method: Agent Mode routes every session to a random model, users work with a single agent in a long single-threaded conversation (sometimes hundreds of turns), the aggregate mixes explicit votes with implicit behavioral signals such as natural-language praise and complaints, and a causal-inference technique called causal tracing — replacing the old Bradley–Terry regression — reports each model’s effect as a contrast in percentage points. That is why the metric is a percentage rather than an ELO.

The statistical cold water: the top three’s confidence intervals overlap. The leader’s lower bound (15.53 − 4.63 = 10.90) sits below third place’s upper bound (13.09 + 2.37 = 15.46), and the No. 1 model has only 4,893 sessions behind it — under 3% of DeepSeek V4.1 Flash’s 182,693. “Who is first” is not a clean statistical claim here. The standing criticism of crowd-sourced evaluation also still applies: blind votes favor longer, more agreeable answers, and Arena itself concedes its vote data is biased and applies statistical corrections. Scoring real task sessions answers part of that critique, but the task distribution remains Arena’s choice. And as of October 12 we found no English-language story dedicated to this week’s board — the weekly write-ups are mostly coming from Chinese tech media.

Our take

For anyone picking a model, two facts on this board matter more than the headline standing. First, the top three’s gaps live inside overlapping confidence intervals, so reading 15.53% against 13.09% as a verdict is overreach; treat the board as a screen and run your own tasks before committing. Second, the cost curve says more than the ranking: $5.41 per task at No. 1 versus $0.13 at No. 17 is roughly a 42x spread for a quarter of the score. For most non-critical work, the “good enough” tier finally has a citable coordinate.

The Chinese cohort has numeric anchors now too: a leading pack at 3.6%–3.8%, Qwen3.8 Max at 2.35%, and roughly a 4x gap to the top. Keep in mind that gap is measured on Arena’s task mix — a Chinese-language task mix would not necessarily preserve it. Honest unknowns: the week number is IT之家’s framing; the board rolls, so every figure here carries an Oct 8 data stamp checked on Oct 12; per-task costs drift; and the video-board claim rests on a single outlet.