MCP-Injector - OpenAI Agents SDK - Proxy Off - Task Completion Quality: leaderboard

MCP-Injector NeurIPS2026 frozen release: 25 single-server MCP-Bench tasks, three seeds, 75 scheduled tasks per model/stack/condition. Four valid source-native rubric metrics; five other metrics remain held for 18 source values outside the stated 0-10 scale. gpt-4o-mini-2024-07-18 judge, five evaluations per run. All-run source policy assigns zero to every failed run; completed runs retain released evaluation values. This is scored task quality, not runner completion. Exact released MXFP4 checkpoints, LM Studio0.3.39, 16000-token context, default decoding. No inferred reasoning effort. 3000s timeout; MCPBench40rounds/OpenAI40turns/LangChain25supersteps. Proxy-off, null pass-through and calibrated18-rule faults remain separate. Native stack=openai; condition=proxy-off; metric=task_completion_score. The released selected campaign is used without own rerun selection; correlated quality subscores share tasks and are not independent experiments. Metric: Mean native task-quality score (0-10).

Metric: Mean native task-quality score (0-10). Source: gitlab.com. Saturation forecast: Rough model projection: around 2026. 2 models tracked.

Top models

Interactive version: theaggregate.ai/benchmark?slug=mcp-injector-openai-agents-sdk-proxy-off-task-completion-quality · How It Works · Data refreshed daily, snapshot 2026-10-27.