GreenPT / greenpt/glm-5.2-caveman-ultra

GLM-5.2 Caveman Ultra - access through LLMTR

Caveman Ultra is GLM-5.2 running under a software-assistant instruction the provider injects on its side. In measurement it described itself as terse and technical with no filler: it drops pleasantries, preserves code, commands, error strings and numbers verbatim, and adds no decorative formatting. That instruction is added server-side and billed to you as input on every request: an eight-token user message was billed as 293 input tokens in measurement. If you want the same model without the extra instruction, use the plain GLM-5.2 row. In measurement the base, lite and ultra variants did not separate: their hidden instructions are within a few tokens of the same length and their answers matched in tone and content. Nothing indicates a quality ladder between them. Because of the provider's injected instructions the model sometimes introduces itself as another company's assistant; the model you are billed for, and the one producing the answer, is GLM-5.2. It offers a 1,000,000-token context window, tool/function calling, and prompt-cache reads billed at a quarter of the normal input rate on repeated prefixes. It accepts text only — no image input. The model reasons by default and those hidden tokens bill as output.