Claude Opus 5.5 takes the lead in Agent Arena, the ranking that distinguishes language models (the systems that generate text) based on their ability to run an autonomous agent, meaning a system that independently chains tasks from an instruction. It relegates Claude Fable 5.1, which was first in September, to third place, with GPT-6.1 Sol slotting in between the two. The platform also publishes its first alignment index, the top of which is entirely dominated by OpenAI.
Available since September 22 and evaluated in Max mode, Claude Opus 5.5 outperforms all models in the ranking, including Claude Fable 5.1, which is priced two and a half times higher via the API, the interface that allows developers to integrate a model into their tools. With Gemini 4 Argon in fourth place, Google achieves the biggest improvement of the month: its best representative, Gemini 3.6 Flash, was ranked 34th a month ago. Claude Sonnet 5.5, sixth, and GPT-6 Sol, seventh, complete the list of new entrants.
All five models already present in September have fallen back. GPT-6 Astra drops from second to fifth place, Claude Opus 5 from third to eighth, Claude Fable 5 from fifth to ninth, and GPT-5.6 Sol from seventh to tenth. Claude Opus 4.8, Kimi K3, Claude Sonnet 5, and GPT-5.5 drop out of the top 10, where Anthropic now fields five models, OpenAI four, and Google one. For each model, Arena publishes a possible rank range, and those of the five followers of Opus 5.5 all include first place. The full ranking is available on the Arena website.
Code, Chat, Work: Claude Opus 5.5 Leads All Three Categories
Arena breaks down its ranking by the type of task assigned to the agent. The Code category includes development, debugging, automation, and data analysis, while Chat covers more personal uses, from creative writing to everyday research. Work groups together office tasks, such as creating documents or professional writing. Claude Opus 5.5 ranks first in all three.
In code, Claude Sonnet 5.5 stands out: third in the category, it gains three places compared to the overall ranking, while Gemini 4 Argon drops out of the top 5. In chat, Gemini 4 Argon climbs to second place, just behind Opus 5.5, and GPT-6.1 Sol, second in the overall ranking, falls to fifth place. The rank range for Opus 5.5 is also wider here than elsewhere, from first to twelfth place. The Work category, which accounted for 69% of tasks at the launch of the categories in mid-August, exactly replicates the top 5 of the overall ranking.
Alignment: OpenAI Occupies the Top Five Spots
Since October 8, Arena also publishes an alignment index, which measures the frequency of three deviations in real sessions: unauthorized action, when the agent exceeds the user’s request or permissions; false attribution, when it attributes statements or choices to the user that the exchanges contradict; and deceptive completion, when it announces a task as finished when it is not.
Among the 27 models evaluated, the five best-aligned are GPT-6.1 Sol, GPT-6 Astra, GPT-6 Luna, GPT-6 Sol, and GPT-5.6 Sol. Claude Opus 5.5, Anthropic’s top model, only ranks sixth, and Gemini 4 Argon eleventh. According to Arena, deceptive completion affects on average 10% of sessions, and up to 48% in code debugging.
Real Sessions Rather Than Blind Duels
Unlike other Arena rankings, the one dedicated to agentic tasks does not have internet users vote between two anonymous responses. It relies on the real use of Agent Mode, the isolated environment where users assign complete tasks to a tooled agent. Each session is randomly assigned to one of the models, which, according to the platform, allows measuring the specific effect of each. Five signals composing the overall score are published in detail: tasks confirmed as successful by the user, the share of positive reactions to complaints, consideration of corrections, reliability of shell commands, and calls to tools absent from the environment.
In the top 10, Claude Opus 5.5 performs best on consideration of corrections, Claude Sonnet 5.5 receives the most validations, and GPT-6.1 Sol shows the best ratio of positive reactions.
