Find benchmarks for the tasks you care about, from coding and research to everyday assistance.
AA-Briefcase evaluates agents producing spreadsheets, presentations, memos and other professional deliverables across 91 tasks. Its combined Elo incorporates rubric performance, analytical quality and presentation. Each task runs independently, and the public Lite example is excluded.
AutomationBench tests agents on realistic business workflows across applications and APIs, covering sales, marketing, operations, support, finance, and HR. It evaluates whether agents can coordinate multi-step work and complete tasks. Zapier’s hosted leaderboard measures strict task completion on a private held-out set.
AutomationBench-AA evaluates agents working across simulated business applications. Its score measures the share of task objectives completed without guardrail violations. This objective-level metric is separate from the strict task-completion score on Zapier’s hosted AutomationBench leaderboard.
GDP.pdf tests reasoning over long professional documents across ten domains. The All-pass score measures attempts satisfying every task criterion. These results use the independent five-attempt evaluation with LiteParse text and page images for models supporting images; other implementations remain separate.
GDPval-AA evaluates models on professional knowledge-work tasks from OpenAI’s GDPval dataset, including creating documents, spreadsheets, slides, and other work products. Agents use tools through the Stirrup framework. Results are Elo ratings from blind pairwise comparisons, rather than percentage accuracy.
LiveBench 2.0 evaluates long-running agentic work on its October 25, 2026 release. This suite contains 18 tasks or questions, scored from 0 to 100 in one run per model. Models use their highest supported effort unless a different setting is reported. This is an individual task-suite score, separate from LiveBench’s weighted overall average.
LiveBench 2.0 evaluates data analysis on its October 25, 2026 release. This suite contains 150 tasks or questions, scored from 0 to 100 in one run per model. Models use their highest supported effort unless a different setting is reported. This is an individual task-suite score, separate from LiveBench’s weighted overall average.
LiveBench 2.0 evaluates document and presentation creation on its October 25, 2026 release. This suite contains 25 tasks or questions, scored from 0 to 100 in one run per model. Models use their highest supported effort unless a different setting is reported. This is an individual task-suite score, separate from LiveBench’s weighted overall average.
LiveBench 2.0 evaluates research reports on its October 25, 2026 release. This suite contains 22 tasks or questions, scored from 0 to 100 in one run per model. Models use their highest supported effort unless a different setting is reported. This is an individual task-suite score, separate from LiveBench’s weighted overall average.
BrowseComp evaluates web research agents on questions with hard-to-find but verifiable answers. It tests persistent browsing, search, and the ability to connect information from multiple sources. Tool access, search budgets, and context management can differ across published comparison reports.
MCP-Atlas evaluates real-world tool use through the Model Context Protocol. It tests tool discovery, choosing correct arguments, recovering from errors, and combining tool outputs to complete multi-step workflows. The stored results below cover the public subset under the April 2026 protocol.
DeepSWE evaluates coding agents on original, long-horizon software engineering tasks. It tests whether an agent can understand a codebase, make changes, and complete engineering work beyond short coding questions.
GPQA Diamond tests graduate-level scientific knowledge and reasoning with difficult multiple-choice questions in biology, chemistry, and physics. Diamond is the 198-question subset selected for high-quality expert agreement. The results here use the independently reported Diamond evaluation.
HLE-Diamond is a revised 1,000-question version of Humanity’s Last Exam, split into 500 reasoning and 500 knowledge questions. These results use the official prompts without tools. Scores are separate from the original Humanity’s Last Exam.
Humanity’s Last Exam evaluates advanced knowledge and reasoning across a broad range of academic subjects using expert-written questions. It includes text and multimodal problems designed to challenge frontier AI systems. Results identify whether the full exam or a text-only subset was used.
LiveBench 2.0 evaluates coding with LiveBench on its October 25, 2026 release. This suite contains 29 tasks or questions, scored from 0 to 100 in one run per model. Models use their highest supported effort unless a different setting is reported. This is an individual task-suite score, separate from LiveBench’s weighted overall average.
LiveBench 2.0 evaluates reasoning with LiveBench on its October 25, 2026 release. This suite contains 202 tasks or questions, scored from 0 to 100 in one run per model. Models use their highest supported effort unless a different setting is reported. This is an individual task-suite score, separate from LiveBench’s weighted overall average.
Terminal-Bench evaluates agents on challenging tasks performed through a terminal. It tests multi-step command-line work, problem solving, and the ability to use tools to complete practical tasks. Reported scores depend on the benchmark release and agent harness.
AppWorld evaluates agents completing everyday tasks across nine simulated apps through code and APIs. Tasks include coordinating messages, shopping, and personal information. Scores here measure Task Goal Completion on the Test-Challenge split: the percentage of tasks for which all checks pass. Agent methods are retained with each result.
IFBench tests whether models follow precise output requirements and generalize to 58 diverse, verifiable constraint types. It measures instruction-following reliability rather than completion of a full agent workflow.
LiveBench 2.0 evaluates browser workflows on its October 25, 2026 release. This suite contains 25 tasks or questions, scored from 0 to 100 in one run per model. Models use their highest supported effort unless a different setting is reported. This is an individual task-suite score, separate from LiveBench’s weighted overall average.
π-Bench evaluates assistants across 100 multi-turn tasks in persistent workspaces. It tests whether an agent notices unstated preferences and requirements, asks useful questions, and carries context across sessions. Scores here use overall Proactivity (PROC), the share of hidden intents resolved proactively. Completeness is a separate metric and is retained with the source results.
Gaia2 evaluates assistants coordinating tasks across simulated email, calendars, shopping, contacts, and other apps. It tests execution, search, adaptation, ambiguity handling, and time constraints. This ranking uses the Gaia2-CLI leaderboard with the OpenClaw 2026.4.1 harness and reports overall pass@1. The older vanilla, noise, and agent-to-agent results are separate evaluations.
LifeVal evaluates personal assistants on everyday responsibilities, including caregiving, household administration, shopping, education, personal finances, and accessing government services. Scores are unweighted average rewards across 169 tasks, expressed as percentages, rather than a strict task-completion rate. The launch results cover five OpenAI models.
MyPCBench evaluates assistants on 184 tasks across a personal desktop with 17 simulated web apps. Tasks involve personal records, travel, shopping, calendars, and coordinating work across applications. These results use the canonical computer-and-bash setup with a 100-step limit and report the percentage of tasks where every rubric item passes.
OSWorld evaluates computer-use agents on practical tasks in a desktop environment. It tests whether agents can interact with applications and complete multi-step work through a graphical interface. Results below are restricted to the stored 2.1 full-set evaluation.
τ²-bench evaluates conversational agents completing customer-service tasks in airline, retail, and telecom environments. Agents use tools, follow policies, and coordinate with a simulated user who can also act. Scores here are overall pass@1 from the Standard leaderboard, using GPT-5.2 as the user simulator. Custom, legacy, voice, and banking evaluations are separate.
