Benchmarking LLM Bidding Agents in Electricity Markets: Hour-Specific Pricing and Experience Effects

LLM bidding agents are an increasingly plausible presence in electricity markets, yet no prior work has calibrated LLM bidding prices against hourly competitive and joint-profit-maximizing references in electricity markets, or examined whether accumulated market exposure shifts that position. We develop a benchmark com- bining a controlled uniform-price market, exact hourly economic references, a normalized price index φ (0 = competitive; 1 = joint-profit maximizing), and a shared-history base–stress–return protocol with a no-stress counterfactual. Agents receive full price history and profit-maximizing instructions with chain-of-thoughts reasoning. Two findings from a five-firm test case stand out. First, the peak-demand hour (H14) reaches φ14 = 0.907, far above the hourly pure-Nash-equilibrium up- per bound of 0.569; the daily average ¯φ = 0.330 blends 24 hourly outcomes and conceals this peak-hour concentration. Second, high-load exposure leaves a persistent gap of +0.068 in daily φ and +0.028 at H14 after conditions normalize; the low-load effect dissipates. hese outcomes do not establish15 collusion or equi- librium convergence; they show that prices produced by AI bidding agents depend jointly on hour-specific pivotality and accumulated market history, both invisible to standard aggregate monitoring.