As of 2026-11-16, Mercor's APEX-Agents leaderboard lists Hy4 High with a Pass@1 of 77.6% ±4.5% and a Mean Score of 85.1% ±3.6%. This is a leaderboard snapshot of Hy4 on professional, multi-step agent tasks in APEX-Agents 1.1; it does not represent general-purpose model capability. The leaderboard does not disclose Hy4's operating costs.
Tasks this benchmark can inform: Agent capability in completing investment banking analysis, management consulting, and corporate legal workflows across applications.
Tasks this benchmark cannot be generalized to: Single-turn Q&A, model capability in isolation, success rates in other tool environments, or general API usage costs.
Applicable model version: The leaderboard identifies the model only as Hy4, with the High reasoning level; it does not specify the server-side build or weights version.
Test environment or client: Mercor APEX-Agents 1.1 and the Archipelago execution environment; the reference agent uses a LiteLLM-based MCP tool-calling loop. The leaderboard does not identify Hy4's specific service provider.
Reasoning level and parameters: High; Hy4's sampling parameters and additional reasoning settings are not specified.
APEX-Agents tests whether model-driven agents can complete long-horizon tasks in a multi-application environment. The three work domains are investment banking analysis, management consulting, and corporate legal work. The leaderboard page says the current benchmark contains 31 worlds, 240 tasks, and scoring rubrics; the APEX-Agents 1.1 update notes say 80 tasks were selected and reviewed in each domain.
Each task is designed by domain experts and includes project files, tools, and an application environment. Outputs are scored against expert-written rubrics and evaluated by an LLM judge. The 1.1 update notes, published on 2026-09-08, describe task reviews for handling ambiguous tasks and broad, scattershot answers, as well as judge updates and system prompt adjustments. The notes say the judge used DeepSeek-v4-Flash-0731 (temperature 0.1) as its baseline and that prompts were optimized using 1,407 rubric annotations. This configuration is for the benchmark judge.
The official reference agent's GitHub documentation says tool outputs are truncated to 200 lines or 32K characters, each run defaults to a maximum of 100 steps and a timeout of 10,800 seconds, and these settings match the leaderboard runs. Reproduction can use Harbor and the public data; some of the full task set remains unpublished.
Metric definitions:
Pass@1: The proportion of tasks for which all rubric criteria are passed in a single run.
Mean Score: The average proportion of rubric criteria passed across all tasks; the leaderboard's primary ranking metric.
Pass@1: Hy4 High 77.6% ±4.5%.
Mean Score: Hy4 High 85.1% ±3.6%.
The leaderboard page does not define the statistical meaning of the ± ranges above. This note retains the values as displayed and does not interpret them as confidence intervals.
The leaderboard does not disclose Hy4's token usage, cost per task or successful task, or API pricing, so cost-effectiveness cannot be compared from these results.
The leaderboard's investment banking analysis example task asks the agent to create an LBO sensitivity table and identify the thresholds at which changes in growth rate, customer acquisition cost, R&D cost, cost of debt, and EBITDA multiple push IRR below 20%. The example page shows a Gemini 4 Carbon sample, not Hy4's per-task result.
| Leaderboard snapshot | Leaderboard name | Model and level | Pass@1 | Mean Score | Cost |
|---|---|---|---|---|---|
| Collected 2026-11-16 | APEX-Agents | Hy4 High | 77.6% ±4.5% | 85.1% ±3.6% | Not disclosed |
The leaderboard is a dynamic page and does not show a separate last-updated date for this row. The date indicates the values shown on the page when this note was collected; it is not Mercor's stated evaluation completion date.
These results are useful for comparing Hy4's performance on Mercor's professional-task agent benchmark. The scores come from an agent working with tools, files, and application environments; they are not an isolated model test outside an execution framework. The public methodology describes the task domains, scoring approach, and some run settings, but the full task set, Hy4's per-task traces, server-side version, detailed sampling parameters, and operating costs are not disclosed in the materials reviewed. The leaderboard score should not be generalized to completion rates across all real-world work scenarios, nor directly compared with results using different task sets or agent harnesses.
On 2026-11-16, this record was collected in Tabbit from the Mercor APEX-Agents leaderboard, switching between Pass@1 and Mean Score and recording the Hy4 High row. Method details were taken from the APEX-Agents 1.1 update notes, the official reference agent and run instructions, the Hugging Face dataset, and the investment banking analysis example page. The APEX-Agents research paper was published in 2026-01, before the 1.1 update; this note does not use the paper's initial sample and scores in place of the current leaderboard's metrics.
Hy4