AI Model Leaderboard for Nonprofits

A ranking of frontier AI models across five pillars weighted for nonprofit priorities: capability, alignment & honesty, cost & efficiency, bias & harm resistance, and security posture.

This leaderboard is an aggregate of independent public benchmarks from reliable sources. It's not based on tests performed by our organization.

The default overall score is calculated with these weights: capability 40%, alignment and honesty 15%, cost and efficiency 15%, bias and harm resistance 15%, and security 15%.

If your organization has different priorities (for example, giving more weight to cost, alignment, bias or security), use the "Change priorities" buttons to re-rank the models, or compare the pillar columns directly.

Expensive models usually require more energy and resources, so higher cost usually translates also as higher environmental impact.

Click a model name to see score details. Click a column header to sort.

Change priorities:

#	Model	Provider	License	Overall score	Capability	Cost	Alignment	Bias & harm	Security

Greener means higher score, redder means lower score. (!) score proxied from the previous model version. (≈) pillar imputed at the set median.

Methodology, sources & limitations

How the overall score works

Each raw metric is normalized to a 0-100 scale with min-max within this tracked set of models (cost per task inverted, so cheaper = higher). Pillar scores are the mean of their normalized metrics. The overall score is a weighted mean of the five pillar scores:

Capability 40%: Artificial Analysis Intelligence Index
Alignment & honesty 15%: Arena.ai Alignment Index (unauthorized action, false attribution, deceptive completion, measured in 72,509 real agent sessions)
Cost & efficiency 15%: Artificial Analysis Cost per Task: the weighted-average USD cost to complete one Intelligence Index task, reflecting actual token consumption (including reasoning verbosity) and caching. List prices are shown as context only.
Bias & harm resistance 15%: Enkrypt red-team ratings: bias, harmful content, toxicity
Security posture 15%: Enkrypt insecure-code and CBRN resistance ratings

The weights are a documented hypothesis derived from nonprofit sector priorities (cost pressure, donor/constituent trust, bias concerns) and are open to challenge. Because every pillar score is published separately, any organization can re-rank the table with its own weights using the open data file. The "Change priorities" buttons above recompute the overall score immediately in your browser (the selected pillar at 40% and the rest at 15%; capability is selected initially, matching the default weights).

Proxied scores (⚠)

When a newly released model is not yet evaluated by a source, its score is inherited from the nearest previous model from the same provider (for example GPT-6.1 Sol inherits GPT-6 Sol's Enkrypt ratings). Every proxied score is marked with ⚠ and a callout naming the source model. Proxied models can outrank fully measured ones.

Imputed pillars (≈)

When no data and no predecessor exist for an entire pillar (e.g. HY4 Preview is not benchmarked by Artificial Analysis at all), that pillar is imputed at the set median so the composite stays comparable. A missing pillar is neither rewarded nor punished. Imputed pillars are marked ≈ and listed per model; models with imputed pillars carry the partial badge, so read their rank with caution.

Known gaps, stated plainly

No live benchmark measures representational fairness (who is depicted, classification parity across groups) for current frontier models. The "bias & harm" pillar measures behavioral resistance: how often adversarial red-team attacks succeed in eliciting biased, toxic or harmful content. That is a different question, and we do not pretend otherwise.
Enkrypt AI is a commercial vendor selling red-teaming and guardrail products, and its prompt set is proprietary. Its scores now carry the bias & harm pillar and part of security. This is a conflict-of-interest risk we disclose openly; a neutral cross-check is a stated goal for v2.
Prices are list API prices, not nonprofit-effective prices. Grants and discounts (Microsoft, Google, OpenAI nonprofit programs) can change the ranking materially; check current programs.
Scores are a snapshot (see date above). Model versions and safety tuning change frequently.
Model name variants across sources (e.g. Enkrypt's glm-5.3-flash vs the GLM 5.3 Max variant scored elsewhere) are reconciled manually and noted in the data file.

Sources

Not an endorsement

This index summarizes third-party measurements for decision support. It is not a safety certification, not legal advice, and not an endorsement of any vendor.