OpenAI-Anthropic cross-eval second round and the permanence of cross-lab safety infrastructure

The second round of OpenAI-Anthropic joint cross-lab safety evaluations establishes cross-lab evaluation as a permanent fixture of frontier-model alignment infrastructure. Combined with the METR cross-lab internal-agent pilot, two-tier cross-lab evaluation is now the operational baseline — a structural achievement of H1 2026.

OpenAI and Anthropic publishing the second round of joint cross-lab safety evaluations is the kind of alignment-infrastructure story that doesn't generate cycle-day headlines but defines the field's operational baseline for the next decade.

From experiment to fixture

The first OpenAI-Anthropic joint evaluation in late 2025 was treated as a proof-of-concept — would two competing frontier labs really run each other's safety evaluations on each other's published models, and would the results be useful enough to justify the coordination overhead? The second round at the 9-month mark answers both questions: yes the labs ran the evaluations, yes the results were useful enough to repeat, and yes the cadence is now treated as a permanent fixture rather than a one-time experiment.

The two-tier cross-lab structure that emerged

The OpenAI-Anthropic bilateral pattern operates at one tier of cross-lab evaluation: direct-bilateral coverage of the two most-capable frontier labs' publicly released models. Yesterday-PM's METR cross-lab pilot covers a different tier: third-party-mediated evaluation of internal-developer agents across four labs (Anthropic, Google, Meta, OpenAI). The two tiers don't overlap; together they cover both the highest-stakes external-deployment models and the highest-stakes internal-tooling deployment surface area.

What this connects to in the Anthropic Fellows pipeline

The Anthropic Fellows Program opening May and July 2026 windows formalizes the six-area alignment-research taxonomy (scalable oversight, adversarial robustness, AI control, model organisms, mech-interp, model welfare) that organizes the broader research field. The cross-lab evaluation infrastructure feeds into this taxonomy at the AI control / scalable oversight tiers — and Anthropic's hiring pattern is increasingly using cross-lab evaluation findings to identify research questions worth funding.

The H2 2026 alignment-infrastructure-maturity moment

Cross-lab evaluation cadence (twice in 9 months) plus METR cross-lab pilot plus Anthropic Fellows six-area taxonomy plus the broader Constitutional AI 2 deployment data produce the most coordinated alignment-infrastructure quarter on record. The H1 2027 alignment-research output should reflect the compounding infrastructure investment — production research findings should land at higher cadence and broader scope than 2024-2025 baseline rates.

Why this matters for procurement

Cross-lab evaluation evidence becomes a default input for high-stakes enterprise vendor selection. Buyers evaluating models for safety-critical deployments will increasingly request cross-lab evaluation findings as part of vendor evidence packages. The H2 2026 vendor-selection-for-safety-critical-workflows pattern increasingly looks like: 'has this vendor's model been through cross-lab evaluation, and what did the findings show?'

OpenAI — Findings from a pilot Anthropic-OpenAI alignment evaluation exercise → · Anthropic — Alignment Research →