Deepseek R1 vs Claude 3 7 Sonnet: $$0.55 vs $$3.00 / 1M Tokens
1. Direct Input & Output Rate Comparison: Deepseek R1 costs $0.55/1M input and $2.19/1M output, while Claude 3 7 Sonnet costs $3.00/1M input and $15.00/1M output. Switching uncached workloads to Deepseek R1 reduces raw input spend by 82%.
2. Prompt Caching & 5-Minute TTL Economics: Deepseek R1 provides 75% automatic disk cache read ($0.14/1M), $0 write fee, whereas Claude 3 7 Sonnet uses 90% explicit cache read ($0.30/1M), 1.25x ($3.75/1M) write fee (5m TTL). Pinning static repository prefixes in Coding Agents reduces realistic 25%-TTL write-retained spend by 61.60% to 65.38%.
3. Context Window & Cliff Thresholds: Deepseek R1 supports 128K tokens and Claude 3 7 Sonnet supports 200K tokens. Pruning sliding-window history below 128K/198K avoids 2x context-cliff surcharges.
4. P90 Streaming Latency (TTFT): In 24-hour rolling probes, Claude 3 7 Sonnet achieves the faster P90 Time-To-First-Token (920ms vs 610ms).