A pre-registered measurement of 11 models on a frozen battery of 60 legitimate prompts. Published as pre-committed: the decay people describe is mostly not visible in refusal, hedging or completeness on this battery. Codes assigned by language-model judges, not validated by humans.
Download the file or read the study for its structure; both are linked above.
Released under CC BY 4.0. You may use it commercially and redistribute it. Attribution is the only condition.
Couey, V. W. (2026). The AI Helpful-Assistant Decay Index 2026: How Often Assistants Refuse, Hedge or Fail to Answer Legitimate Requests. Deep Synthesis. https://doi.org/10.5281/zenodo.23730212
Found an error, need a different cut of this data, or citing it in something? vincent@deepsynthesis.org. Corrections are made against the published file and the DOI is versioned, so a fix is traceable rather than silent.
Time-series of input/output token prices for 12 frontier model families across 32 versions, with cost-per-reference-task normalized per entry.
Per-tool pricing matrix across the Amazon seller SaaS ecosystem (Helium 10, Jungle Scout, Keepa, et al.) with feature-tier delta and effective monthly cost.
Per-claim evidence audit of common standing-desk health and productivity marketing claims, sourced to peer-reviewed and meta-analytic literature.
Pricing and capability index of 30 go-to-market AI tools across vertical and job-to-be-done, with entry monthly cost, pricing model, and verified-source date.