How is Claude priced?

Per token, billed separately for input and output, with output priced higher. Rates differ by model tier — Haiku is the cheapest, Opus the most expensive.

Do caching and batching change the price?

Yes. Cached input tokens and batch processing are billed at reduced rates, which can materially change the effective cost of a repetitive workload.

Which Claude model should I budget for?

Price the same workload on Haiku, Sonnet and Opus and compare. Many production tasks run acceptably on the mid tier, and the cost gap between tiers is usually larger than the quality gap.

Anthropic Claude Pricing: Sonnet, Haiku, and Opus Rates

Current per-token rates

Claude 3.5 Sonnet: $3/$15 per 1M. Claude 3.5 Haiku: $0.80/$4. Claude 3 Opus: $15/$75. All billed per token.

Prompt caching economics

Cache write: 1.25× standard input rate. Cache read: 0.1× standard input rate. Break-even at ~2 reuses. High-reuse system prompts get 90% off input.

Batch API and other discounts

Anthropic Batch: 50% off both sides. Enterprise contracts available above ~$20K/mo with custom rate cards.

Worked cost example

RAG assistant: 8K context + 600 output, 50K calls/month, 70% cache hit rate. Sonnet uncached ≈ $1,650. With caching ≈ $620. Caching pays back in 24h of traffic.

When each tier fits

Sonnet: default for production reasoning. Haiku: chat, classification, high-volume routes. Opus: research and edge cases only — Sonnet now matches it on most benchmarks at one-fifth the price.

Mechanics: why cache writes cost more than standard input

Writing to Anthropic's prompt cache costs 1.25x the standard input rate because the system has to process and store the prefix for later reuse, an extra step beyond a normal pass. That premium is why caching only pays off when a prefix is reused at least twice — a one-off cached call costs more than not caching at all, since you paid the write premium with no read-side saving to offset it.

Decision checklist for choosing a Claude tier

— Is the system prompt or context reused across many calls? Cache it. — Does the task need frontier reasoning, or does Sonnet already clear your accuracy bar? — Is the workload latency-tolerant enough for the 24h Batch SLA? — Have you priced Haiku for the high-volume, low-complexity slice of traffic separately?

How to sanity-check a Claude invoice

Back out your blended $/1M by dividing invoice total by total tokens logged; compare it against $3/$15 (Sonnet) or $0.80/$4 (Haiku) weighted by your actual model mix and cache-hit rate. If your invoice implies a lower rate than the raw list price, that's consistent with a working cache; if it's higher, check whether cache breakpoints were actually reached (they require exact prefix matches ≥ Anthropic's minimum cacheable length).

Trade-offs this pricing structure creates

— Cache TTLs (typically 5 minutes) mean bursty or low-frequency traffic gets none of the 90% discount. — Opus remains 5x Sonnet's rate, so a routing bug sending traffic to Opus is expensive to miss. — Batch's 24h window rules it out for anything synchronous, regardless of the discount size.

Frequently asked questions

How is Claude priced?: Per token, billed separately for input and output, with output priced higher. Rates differ by model tier — Haiku is the cheapest, Opus the most expensive.
Do caching and batching change the price?: Yes. Cached input tokens and batch processing are billed at reduced rates, which can materially change the effective cost of a repetitive workload.
Which Claude model should I budget for?: Price the same workload on Haiku, Sonnet and Opus and compare. Many production tasks run acceptably on the mid tier, and the cost gap between tiers is usually larger than the quality gap.