Claude Opus 5.5 Claims #1 on Agent Arena as Muse Wonder and Haiku 5.5 Make Their Debuts
Arena’s Agent Arena leaderboard updated on October 14 with 2.58 million total sessions across 54 models. Three changes stand out: Claude Opus 5.5 reached #1, Meta’s Muse Wonder made its first appearance at #12, and Claude Haiku 5.5 entered at #27.
Claude Opus 5.5 at the Top
Claude Opus 5.5 Max leads with a 17.18% net improvement rate, 19.36% confirmed success, 17.94% steerability, and 15.50% bash reliability. It rose 16 positions from rank 17, accumulating those metrics across 6,796 sessions.
That session count is low relative to older models on the board. Claude Opus 5 at rank 8 has 23,614 sessions; GPT-5.5 at rank 33 has 82,556. Opus 5.5 reached #1 on a narrower sample, which typically indicates high per-session performance on the harder, longer agent tasks that experienced Arena users submit.
At $5.35 per completed task (P50), it is the second-most expensive path at the top behind Fable 5.1 at $7.23.
The Cost Gap Among the Top 5
The five models within 3.5 percentage points of the top show a wide spread in cost per completed task:
| Rank | Model | Net Improvement | Cost/Task |
|---|---|---|---|
| 1 | Claude Opus 5.5 (Max) | 17.18% | $5.35 |
| 2 | GPT 6.1 Sol (Max) | 14.78% | $0.65 |
| 3 | Claude Fable 5.1 (Max) | 14.12% | $7.23 |
| 4 | Gemini 4 Argon (High) | 13.73% | $0.85 |
| 5 | GPT 6 Astra (Max) | 13.69% | $3.34 |
GPT 6.1 Sol at rank 2 costs $0.65 per task — 8x less than Opus 5.5 and 11x less than Fable 5.1. Gemini 4 Argon at rank 4 costs $0.85. The performance-per-dollar case for both is difficult to dismiss given only a 3.5-point net improvement gap between Sol and the top spot.
Argon also records the lowest tool hallucination rate of any model in the top 20: 0.02% across 28,097 sessions. In production agent deployments, hallucinated tool calls generate failed API requests and broken automations. At rank 4, Argon delivers both performance and tool reliability at $0.85/task.
Muse Wonder Debuts at #12
Meta’s Muse Wonder (Max) entered Agent Arena at #12 with 5,526 sessions. Its figures: 6.42% net improvement, 8.43% confirmed success, 8.68% steerability, and 0.27% tool hallucination. The hallucination rate is the lowest of any model joining the board this week and below the top-5 average of 0.48%.
Confirmed success at 8.43% compares reasonably to GPT-6 Sol at rank 7 (9.21%) despite Muse Wonder sitting 5 places lower overall. The $1.31 cost per task lands between the cheapest Google option (Gemini 3.8 Flash at $0.72, rank 14) and GPT-6 Astra ($3.34, rank 5), making it competitive on cost for mid-tier agent work.
As of October 15, The Agent Arena debut follows Muse Wonder Full’s recent showing on Artificial Analysis’s Intelligence Index, where it placed among the top five models. The task-completion data from Agent Arena is consistent with that general capability positioning.
Claude Haiku 5.5 Enters at #27
Claude Haiku 5.5 High ranked 27th on its first day, posting 1.88% net improvement, 6.40% steerability, and 7.78% bash reliability across 21,839 sessions. That session count exceeds most models ranked above it — including Claude Opus 5 (23,614), Fable 5.1 (16,282), and GPT-6 Astra (12,892).
The driver is cost: $0.10 per completed task. Two models on the 54-model board undercut it on price: MiMo V2.6 Flash at $0.05/task (rank 22) and GPT-6 Luna at $0.08/task (rank 19). Haiku 5.5’s predecessor, Haiku 4.5, scored 0.0% on Terminal-Bench 4.0. Haiku 5.5 scored 39.2% on the same benchmark, and its 7.78% bash reliability on Agent Arena reflects that improvement in real sessions. The combination of low price and credible bash performance makes it a practical subagent tier for pipelines that previously required Sonnet-class spending.
Tool Hallucination as a Production Signal
One outlier in the full 54-model leaderboard: Mistral Large 4 at rank 39 with 17.61% tool hallucination across 14,022 sessions. That is roughly 1 in 6 tool calls hallucinated, compared to Argon at 0.02% and Muse Wonder at 0.27%. Mistral Large 4 launched in early October with strong scores on finance and legal vertical benchmarks; Agent Arena is exposing a reliability gap in unstructured agentic contexts.
What Shifted
On October 13, a batch including GPT-6 Astra, GPT-6.1 Sol, Claude Fable 5.1, and Claude Opus 5.5 entered Arena’s Search Arena leaderboard. On October 14, Haiku 5.5 and Muse Wonder joined Agent Arena specifically. The two-day expansion brings the total to 54 agent-evaluated models, with Anthropic holding seven of the top 13 positions and OpenAI holding four.
The net improvement gap between rank 1 (17.18%) and rank 5 (13.69%) is 3.5 percentage points. The cost spread across those five positions runs from $0.65 (Sol, rank 2) to $7.23 (Fable 5.1, rank 3) — an 11x range within positions that sit within 3.5 points of each other on the key outcome metric. That is the current agentic deployment calculus: the top five perform within a tight band on task success while spreading across an order of magnitude on cost.