Frontier AI Labs Comparison

Read this before the table

Epistemic status. Opinionated synthesis, not a peer-reviewed scorecard. Ratings reflect publicly observable signals through early 2026; internal practices that aren't in the public record are not assessed.

Direction of ratings. For every safety dimension, HIGH means more safety-positive (more investment, more restraint, more transparency, stronger alignment, more detail). The capability "Tier" column is descriptive, not evaluative — being a frontier lab is not in itself good or bad. UNCLEAR means we don't have enough public signal to rate — not that the underlying practice is bad.

Limits. Ratings age fast — frontier labs reorganize, ship new policies, and lose senior staff on quarterly timescales. Update cadence here is ~14 days; always check the linked references for current state. We weigh observable artifacts (published RSPs, system cards, attrition patterns) more heavily than press statements. Other scorecards weight differently — see FLI AI Safety Index, METR Common Elements, and AI Lab Watch for independent rubrics.

What this table is — and isn't

It's a cross-lab comparison overlay on top of the FactBase / TableBase organization data this wiki already maintains. Hard numbers (funding, valuations, headcount) live in the per-org pages at /organizations. This table is the layer above that — the editorial judgment about safety practice, anchored where possible in observable artifacts.

It is not:

• A precise scoring system. The HIGH / MEDIUM / LOW levels are buckets, not measurements.
• A prediction. We rate revealed behavior to date, not where each lab is heading.
• A safety endorsement for HIGH-rated labs. A lab can score well on every column here and still be a net contributor to risk (because the underlying activity is high-risk, or because revealed safety practice can degrade fast).
• Definitive. Other lab-scorecard efforts (FLI, METR, AI Lab Watch) use different rubrics and reach different conclusions on the same labs. We cite ours, not theirs.

Why these eight dimensions

The dimensions are chosen to track what a third-party observer can verify from public artifacts, on the theory that the gap between stated safety commitment and revealed safety behavior is more useful than either alone:

• Capability tier — descriptive, not evaluative. Distinguishes labs whose safety choices actually matter at the global frontier from those whose choices matter mostly to their customers.
• Safety investment — dedicated alignment / interpretability / red-team headcount and research output, relative to org size. Absolute headcount favors hyperscalers; ratio is more meaningful.
• Race posture restraint — public framing plus revealed behavior on competitive deployment. We mark HIGH when restraint is both stated and demonstrated; LOW when racing is explicit or when revealed cadence contradicts safety framing.
• Transparency — RSPs, system cards, training detail, AISI pre-deployment access. Weighted toward operationally specific artifacts (capability thresholds, tripwires) over PR commitments.
• Governance engagement — constructive engagement with regulators, AISIs, summits. We don't credit a lab for testifying if they also lobbied against the same legislation.
• Leadership safety alignment — CEO / co-founder public stance plus revealed prioritization (resourcing, who reports to whom, what gets shipped when leadership disagrees).
• Employee safety culture — internal safety advocacy, public alignment-research output, attrition patterns. High-profile safety-driven departures are weighed heavily because they're a costly signal.
• RSP / framework quality — operational specificity. We treat a vague public-relations document as worse than no policy, because the former forecloses scrutiny.

What HIGH means

HIGH is safety-positive, uniformly. HIGH safety investment means more investment; HIGH transparency means more disclosure; HIGH race restraint means more restraint; HIGH leadership alignment means stronger alignment. This lets you scan the table left-to-right and read green/amber/red as a rough safety signal — at the cost of forcing some unidimensional judgments onto multidimensional realities. The per-cell notes try to recover the nuance.

We use UNCLEAR rather than guessing when public signal is thin (especially for Chinese-jurisdiction labs and very small / very new labs). Treat UNCLEAR rows as "we don't know," not as "probably bad."

Cross-references

• Anthropic Stakeholders Table — single-org-deep-dive prototype.
• /organizations — organization directory with FactBase facts (funding rounds, valuations, headcount).
• FLI AI Safety Index (Winter 2025) — independent scorecard, different rubric.
• METR: Common Elements — frontier safety policy comparison.
• AI Lab Watch — third-party scorecard with detailed rubric.

Methodology — how each rating is anchored

Every rating in the table has a one- or two-sentence note linking it to an observable signal. Where multiple signals exist, the "Notable signals" column carries the longer evidence list and the "References" column links sources.

When evidence conflicts (e.g., a public safety statement contradicted by revealed deployment cadence), we weight revealed behavior over stated commitments, on the theory that the costly-signal side carries more information. This is why labs that publish strong-sounding safety frameworks can still receive a MEDIUM or LOW rating on race-posture restraint — the frameworks are real, but so is the deployment cadence.

For attribution-light claims (e.g., "small explicit safety team"), the absence of public confirmation is itself informative: frontier labs that invest heavily in safety typically advertise it. We mark these LOW, not UNCLEAR.

Limitations

• Sample is small — 12 labs, eight dimensions, hand-curated. Per-cell error bars are wide.
• Western-centric public signal — Chinese-jurisdiction labs operate under a different regulatory and disclosure regime; UNCLEAR ratings there reflect our limited visibility, not necessarily their internal practice.
• Race-posture restraint is the most contested column — labs and observers disagree about what counts as racing. We use "race-conscious" as MEDIUM and reserve LOW for labs whose public framing explicitly endorses racing.
• Ratings age fast — quarterly cadence at minimum. Anything older than three months should be re-verified before being cited.

Suggested re-reads, in order of likely turnover: leadership safety alignment > race-posture restraint > RSP quality > safety investment > transparency > everything else.