Rates, Haiku 4.5: {'inp': 1.0, 'w5m': 1.25, 'w1h': 2.0, 'hit': 0.1, 'out': 5.0, 'b_in': 0.5, 'b_out': 2.5} Example 1. Classification, 10,000 requests x 400 input + 20 output tokens input tokens = 10,000 x 400 = 4,000,000 output tokens = 10,000 x 20 = 200,000 input cost = 4,000,000 / 1M x $1.00 = $4.00 output cost = 200,000 / 1M x $5.00 = $1.00 total standard = $5.00 per request = $0.0005 total on Batch API = 4,000,000/1M x $0.50 + 200,000/1M x $2.50 = $2.50 same job on Sonnet 5 = 4,000,000/1M x $2.00 + 200,000/1M x $10.00 = $10.00 same job on Sonnet 4.6 = 4,000,000/1M x $3.00 + 200,000/1M x $15.00 = $15.00 same job on Haiku 3.5 = 4,000,000/1M x $0.80 + 200,000/1M x $4.00 = $4.00 Example 2. Chat product, 2,000,000 input + 500,000 output tokens per day, 30 days no cache: daily input = 2,000,000/1M x $1.00 = $2.00 no cache: daily output = 500,000/1M x $5.00 = $2.50 no cache: daily total = $4.50; 30 days = $135.00 cache: cacheable 1,500,000/day at 90% hits -> hits 1,350,000, misses 150,000, never cached 500,000 cache hits = 1,350,000/1M x $0.10 = $0.14 cache misses = 150,000/1M x $1.00 (base input) = $0.15 never cached = 500,000/1M x $1.00 = $0.50 output = $2.50 cache: daily total = $3.29; 30 days = $98.55 saving vs no cache over 30 days = $36.45 (27.0%) note: if every miss is re-written as a 5m cache write at $1.25 instead of $1.00, add $0.04/day = $1.12/30 days, total $99.68 note: a single one-time 5m write of a 1,500,000 token prefix = $1.88 same workload, no cache, Sonnet 5: 30 x (2.0 x $2.00 + 0.5 x $10.00) = $270.00 same workload, no cache, Sonnet 4.6: 30 x (2.0 x $3.00 + 0.5 x $15.00) = $405.00 Example 3. Same 30 day chat workload on Batch API rates ($0.50 in, $2.50 out; cache hit 0.1x of batch input = $0.05) batch, no cache: daily = 2.0 x $0.50 + 0.5 x $2.50 = $2.25; 30 days = $67.50 batch + cache: daily = 500,000/1M x $0.50 + 150,000/1M x $0.50 + 1,350,000/1M x $0.05 + 0.5 x $2.50 = $1.64; 30 days = $49.28 batch vs standard, no cache, 30 days: $135.00 -> $67.50 batch + cache vs standard no cache, 30 days: $135.00 -> $49.28 (63.5% lower) Cross-check ratios Sonnet 5 / Haiku 4.5 input 2x, output 2x Sonnet 4.6 / Haiku 4.5 input 3x, output 3x Haiku 4.5 / Haiku 3.5 input 1.25x, output 1.25x 5m write break-even: write 1.25x, hit 0.1x; 1.25 + 0.1 = 1.35 < 2.0 so caching pays after one hit