ToltIQ published its model selection method and benchmark results for the models offered in its Chats feature: Opus 5.5 (92.3% accuracy), Sonnet 5.5 (86.7%) and GPT-6.1 Sol (84.9%) on its CASE Bench of 250 credit agreement questions run three times, plus cross document synthesis and file generation tests. It states models must meet zero data retention terms before being offered, and names models it excludes (Fable 5.1, Mythos 5.1, GPT-6 Astra) and why.
Our readGives a model risk reviewer a named list of third party models in production, the gate each must pass, and run to run accuracy and latency figures to cite in validation work. Results are self reported on a vendor built benchmark and not independently reproduced.