Bedrock Cost Comparison: Claude vs. Llama vs. Nova for the Same Task

All three providers on Bedrock quote per million tokens -- but comparing just that number, without considering cached tokens, batch pricing, and provisioned throughput, can have you paying 2-3x more than necessary for the same work.

Base prices, family by family

Amazon Nova Micro is priced at $0.035 per million input tokens and $0.14 for output; Nova Lite at $0.06/$0.24; Nova Pro at $0.80/$3.20. Meta Llama 4 8B costs $0.18 input / $0.24 output per million tokens; Llama 4 70B, $0.65/$0.86; Llama 4 405B, $1.95/$2.56. Claude Sonnet 5 on Bedrock is priced at $2/$10 per million input/output tokens through August 31, 2026 as a promotional rate, rising to $3/$15 afterward.

The discount almost nobody takes advantage of: cached tokens

Cached input tokens cost roughly 10% of the base rate -- for architectures with a long system prompt or repeated context across calls (like an agent that resends conversation history on every turn), enabling caching can cut the effective cost of that portion of the prompt by 90%, without changing anything about the model's behavior at all.

Batch pricing: 50% off if you can wait

Batch pricing offers a 50% discount over standard rates -- the condition is that the job doesn't need a real-time response. For tasks like mass classification of a historical dataset, overnight report generation, or any pipeline not on the critical path of a user interaction, batch pricing is, literally, half the cost for the same work.

Provisioned throughput: savings in exchange for commitment

Provisioned throughput offers between 15% and 30% savings on predictable, high-volume workloads, in exchange for a one- or six-month commitment -- it makes sense once you already have a stable, known usage pattern, not during the experimentation phase where volume is still uncertain.

How to build the real comparison for your use case

The "$X per million tokens" number on the pricing page is rarely the actual cost you'll pay -- an honest comparison needs to simulate your specific application's real usage pattern (how much context repeats and could be cached? does the work tolerate batch? is volume predictable?) before deciding which model family is cheaper for your case, instead of comparing only list price.