CAISI Evaluates DeepSeek V4 Pro Against Frontier Models

NIST's Center for AI Standards and Innovation published CAISI Evaluation of DeepSeek V4 Pro on May 1, 2026. The evaluation says DeepSeek V4 is the most capable PRC AI model CAISI has evaluated to date. At the same time, CAISI estimates that the model's aggregate capability lags leading frontier models by roughly eight months.

CAISI evaluated DeepSeek V4 across domains including cyber, software engineering, natural sciences, abstract reasoning, and mathematics. The report says DeepSeek V4 scores better on DeepSeek's self-reported evaluations than on CAISI's own evaluations. CAISI's benchmark suite includes non-public evaluations, which is important because public benchmarks can be affected by benchmark leakage, overfitting, prompt tuning, or selective reporting.

The evaluation also discusses cost efficiency. CAISI reports that DeepSeek V4 was more cost efficient than a U.S. reference model on most of the evaluated cost comparisons, while still trailing the frontier in aggregate capability. That creates a more realistic tradeoff for teams: the model may not be the strongest available option, but cost efficiency can still make it competitive for specific workloads.

This is a useful example of why model selection should include independent and private evaluations. Public model cards and release claims are useful starting points, but production teams need tests that reflect their own risk profile, data distribution, latency budget, and failure tolerance. A model that performs well on public coding or reasoning benchmarks may behave differently on internal tasks, adversarial prompts, or domain-specific workflows.

Because CAISI is a public-sector evaluation body, its findings can influence government procurement, enterprise risk reviews, and model-governance discussions. The report does not simply rank models; it demonstrates how independent evaluation can challenge vendor-reported capability narratives while still acknowledging cost advantages.

Watch whether DeepSeek responds with updated methodology, whether other independent evaluators reproduce the eight-month gap estimate, and whether enterprises treat cost efficiency as enough to justify deployment in lower-risk workloads. Practitioners should also watch how future CAISI reports handle open-weight models, frontier closed models, and domain-specific safety evaluations.