Finding / eve

2026-10-13-eve-defaults-that-flipped-an-unconfigured-agent-now-runs-gpt-6-luna-fast-at-high-reasoning-and

Defaults that flipped: an unconfigured agent now runs GPT-6 Luna Fast at high reasoning, and Eve attempts 100% trace sampling on newly created Vercel projects. Unconfigured agents ran GPT-5.6 Luna Fast at 0.63.0, Grok 4.7 from 0.64.0, then GPT-6 Luna Fast with high reasoning from 0.69.0. This changes the default route and reasoning settings, not a measured performance or cost result. Eve deploy (0.66.2), and link/integration project creation from 0.75.1, attempt 100% trace sampling on newly created Vercel projects. The helper catches setup failures and warns; existing projects retain their settings. The opt-out is documented for eve deploy --no-trace-sampling, not every creation command. The runtime audience-to-content table stayed unchanged: development content for any audience, preview/production content only for public conversations. Datadog eval export has a separate, changed default (section 16).

Channel: tagged-release. Half: both. Date: 2026-09-22 (eve@0.64.0), 2026-09-24 (eve@0.66.2), 2026-10-01 (eve@0.69.0), 2026-10-08 (eve@0.75.1).

Operator consequence: Pin model and reasoning when cost, latency or eval baselines matter, and rerun evals after changing the route. For newly created Vercel projects, inspect actual sampling settings and setup warnings instead of assuming every trace was sampled. Review runtime trace policy separately from eval reporter content options.

Receipt

Finding metadata

Run: 2026-10-13-profile-refresh-2026-09-21_2026-10-13-frontier-v0

Finding ID: 2026-10-13-eve-defaults-that-flipped-an-unconfigured-agent-now-runs-gpt-6-luna-fast-at-high-reasoning-and

Profile citations

  • Eve / claim / model-and-trace-setup-defaults
  • Eve / posture / governance

Source links

Primary links, including exact changelog lines when available.