Arena reports that Gemini 4 Hydrogen (High) ranked fifth in the Agent Arena launch results, with a net improvement rate of +14.8% across 6.5k real-world agent sessions. The public post supports a description of that leaderboard snapshot, but does not provide enough information to reproduce the test or infer performance on other tasks.
Tasks this can help assess: Agent task performance and latency snapshots in Arena Agent Arena.
Tasks this should not be generalized to: Other benchmarks, prompts, clients, reasoning tiers, or undisclosed task distributions.
Applicable model version: Gemini 4 Hydrogen (High).
Test environment or client: Arena Agent Arena; Agent Mode tasks. The post does not specify the client or full test configuration.
Reasoning tier and parameters: High; other sampling parameters are not specified.
Arena says the results are based on 6.5k real-world agent sessions. The post defines one model API call during an Agent Mode task as one step. Latency is measured from the first model request through the final streamed response and excludes tool execution. The post does not provide task sampling details, prompts, model-call parameters, or full scoring details.
Ranked fifth in the Agent Arena launch results, with a net improvement rate of +14.8%.
Lowest median task latency among the top 10 models: 5.3 minutes per task. Arena says this is about one-third of Claude Fable 5.2 (Max) and about 1/2.4 of GPT-6 Astra (Max).
Typical latency per model step: 7.3 seconds. Arena says this is about 2.7 times faster than Claude Sonnet 5.5 and 3.1 times faster than GPT-6.1 Sol.
82% of tasks took under 10 minutes, compared with 55% for GPT-6.1 Sol and 44% for Claude Sonnet 5.5. 97% took under 30 minutes, compared with 81% and 78%, respectively.
The data comes from Arena's original post and a benchmark reply in the same thread. The original post does not list confidence intervals, sample sizes for each model, or per-task records; the tiers for comparison models are also not fully specified in the speed post.
This source is Arena's first-party publication of its leaderboard results and is suitable as an Agent Arena snapshot as of 2026-11-13. The post does not explain how the net improvement rate is calculated, the task composition, or statistical uncertainty. Latency excludes tool execution, so it is not end-to-end task latency; the sample size and aggregate results also do not establish that Hydrogen leads on all agent tasks.
This benchmark cannot be reproduced from the public posts alone. Reproduction would require Arena's task set, sampling and scoring rules, model/API configuration, per-task records, and statistical methods.
Gemini 4 Hydrogen