Found first: a primary source the press has not covered yet.
An open harness framework called HEELO has scored 55.5% on Terminal-Bench 4.0 using Dev-Primitives, above Claude Opus 5 with Claude Code at 51.8% and GPT-5.6 Sol with Codex at 37.3%. The paper, "The Last Mile to the Real World: Agent-Native Primitives as Reusable Building Blocks for Solving Real-World Tasks", was submitted on November 5, 2026 by lead author Haibo Jin and 13 co-authors.
What the source says
HEELO (Harness Engineering around Evolving Agent-Native Primitives for the Last Mile to Real-World Tasks) organizes tools into eight families of Agent-Native Primitives, structured as reusable building blocks for real-world task completion. The Dev-Primitives configuration produces the Terminal-Bench 4.0 result: 55.5%, placing 3.7 percentage points above Claude Opus 5 with Claude Code and 18.2 above GPT-5.6 Sol with Codex. The pattern holds on other benchmarks: ToolBench moved from 55.8% to 75.1%, GAIA-Text-103 from 45.6% to 58.3%. On a 46-task real-world platform at heelo.dev, Claude Opus 5.5 and GPT-6.1 Sol with Codex each completed 9 tasks independently; with HEELO, those figures rose to 39 and 35 respectively, with task time reduced by 1.9x to 2.3x and cost per completed task by 1.55x.
Why it matters
The result challenges the assumption that frontier-model capability is the primary lever for agentic benchmark performance. A harness-level intervention, without changing the underlying model, moves the score above the current top commercial agent stack on Terminal-Bench 4.0. The gains across ToolBench and GAIA alongside Terminal-Bench suggest the primitives are not tuned to a single benchmark. The real-world task counts, 39 and 35 completions versus 9 each working alone, are the more direct argument: the paper frames the gap between benchmark scores and deployed agents as partly an engineering problem, and these numbers are its evidence.