diff --git a/src/PowderCoating.Infrastructure/Services/AiCatalogPriceCheckService.cs b/src/PowderCoating.Infrastructure/Services/AiCatalogPriceCheckService.cs
index c2d75fa..783ea96 100644
--- a/src/PowderCoating.Infrastructure/Services/AiCatalogPriceCheckService.cs
+++ b/src/PowderCoating.Infrastructure/Services/AiCatalogPriceCheckService.cs
@@ -22,8 +22,9 @@ public class AiCatalogPriceCheckService : IAiCatalogPriceCheckService
 
     private const string Model = "claude-haiku-4-5-20251001";
     private const int BatchSize = 25;
-    private const int MaxConcurrentBatches = 2; // 3 concurrent bursts past Haiku's output TPM limit
-    private const int RateLimitRetrySeconds = 65; // wait just past the 60s window before retrying a 429
+    private const int MaxConcurrentBatches = 2;
+    private const int RateLimitRetrySeconds = 65;
+    private const int MinBatchIntervalSeconds = 20; // proactive pacing: ~3 batches/min × ~2k tokens = ~6k TPM, under the 8k limit
 
     private static readonly JsonSerializerOptions JsonOpts = new() { PropertyNameCaseInsensitive = true };
 
@@ -137,7 +138,13 @@ public class AiCatalogPriceCheckService : IAiCatalogPriceCheckService
             {
                 _logger.LogInformation("Starting price check batch {Index}/{Total} ({Count} items)",
                     index + 1, batches.Count, batch.Count);
-                return await AnalyzeBatchAsync(client, systemPrompt, batch);
+                var sw = System.Diagnostics.Stopwatch.StartNew();
+                var result = await AnalyzeBatchAsync(client, systemPrompt, batch);
+                // Pace output token rate: hold the slot until MinBatchIntervalSeconds has elapsed
+                // so we stay under the per-minute output token limit without relying solely on retries.
+                var pad = (int)(MinBatchIntervalSeconds * 1000 - sw.ElapsedMilliseconds);
+                if (pad > 0) await Task.Delay(pad, cancellationToken);
+                return result;
             }
             finally
             {