Desciption This preprint introduces Conversational Thermodynamics (CT), a physics-inspired framework for evaluating large language models not by benchmark scores, but by how their reasoning degrades under recursive pressure. CT treats meaning as semantic energy and models conversation as a thermodynamic process governed by entropy, drift, dissipation, and collapse velocity. Using a 250,000-word naturalistic corpus of multi-model human–AI dialogues originally collected for the AiSEON Engagement Analysis Framework (EAF), the study analyzes recursive reasoning, adversarial reframing, ethical dilemmas, and high-load abstraction tasks. CT formalizes three key operators—semantic entropy (Hs), drift rate (δ), and semantic integrity (SI)—and derives a collapse-velocity law, Vc = δ × (1 – SI), that predicts when conversational coherence will fail. Contextual Loops (CTL) and a Semantic Progression (SP) rubric distinguish genuine epistemic traction from fluent stagnation, while the Principle of Recursive Impactrum (PRI), the Principle of Irreducible Instability (PII), and Synaptic Drift Theory explain why instability accelerates with depth in all current large language model architectures. Together, these components allow CT to benchmark AI “fluid intelligence” as the capacity to sustain coherence, adapt under perturbation, and resist drift over recursive cycles. The results support a thermodynamic verdict: present-day predictive architectures cannot self-repair, cannot maintain stable semantic identity across depth, and exhibit an irreducible probability of collapse P(c) > 0. CT therefore reframes AI evaluation from capability scores to the physics of meaning, providing a transferable method for researchers and enterprises to test their own models using their own data.