Tokens Are Not Dollars: Latent vs Explicit Reasoning Inside LLM Agents, Measured in Cost per Solved Task

Latent reasoning is often motivated by a reduction in visible reasoning tokens, but an agent is not paid for by the token alone: failed attempts consume inference, latent steps consume forward passes, and wall-clock latency remains a deployment cost. This paper extends the cost-per-solved-task accounting of *Token-Economics for Agentic AI* to explicit chain-of-thought (CoT), training-free latent feedback, compute-matched controls, and a Coconut-style trained latent checkpoint. In the training-free regime, using Qwen2.5-1.5B-Instruct with 16 latent steps at seed 0, latent feedback solved only 0.014 of both GSM8K and ProsQA tasks. On ProsQA it cost \$0.008012 per solve, compared with \$0.000380 for the matched explicit-CoT seed; the pause-token control matched explicit accuracy at 0.202 and cost \$0.000316 per solve. Adding Gaussian noise to the latent states produced no ProsQA solves. A pre-specified sanity gate then made the trained comparison interpretable: a stage-0 GPT-2 explicit checkpoint reached 0.570 accuracy on ProsQA. On the final Coconut-style checkpoint, explicit CoT reached 0.768 accuracy at \$0.000054 per solve, whereas latent inference with three latent steps reached 0.676 at \$0.000064 per solve, with 2.12× the median latency and 1.18× the FLOPs per solve. Training therefore repaired the catastrophic training-free failure mode, but it did not make latent reasoning cheaper per solved task in the measured setting. The results support cost per solved task—not visible token count—as the decision metric for reasoning-mode selection.