Swfte AI Research Reports — May 2026

Most published model reviews are either marketing reposts or vibes-driven Twitter takes. Neither is useful when a procurement decision is on the line. We built the Swfte Model Quality Test Suite (SMQTS) to fix that — three series of structured prompts (programming, non-programming, cost-quality validation) graded blind across multiple raters with a published 0-5 rubric.

Every model in this collection has been through the full SMQTS at its pinned May 2026 version. The reports below show the category-level scores, the failure modes we found, and the specific workloads where each model wins or loses. We do not award points for marketing copy.

We disclose the obvious conflict: Swfte Connect routes production traffic across these same models, so we have a commercial incentive in the routing decisions readers make. The mitigation is transparency — we publish the prompts, the transcripts, the scoring sheets, and the disagreements between raters. If our scoring is wrong, the raw data lets you prove it.

These reports are written for three audiences: procurement teams evaluating multi-year contracts, engineering leads picking a default model for a product line, and researchers comparing released systems on a common rubric. Each section flags which audience it serves.