Claude API Pricing: Complete Cost Breakdown for Every Model

Claude API pricing follows a pay-per-token model. You pay for what you use, with no minimum commitment. Every model in the Claude family has a different rate, and several built-in discounts can cut your bill by 50% or more.

This guide covers the exact per-token cost for each Claude model as of July 2026. It also explains prompt caching, the Batch API, extended thinking charges, and how the flat-rate Claude Code plans compare to raw API access.

Whether you are building a prototype or running production workloads, understanding these numbers helps you pick the right model and optimize your spend.

Claude API Pricing by Model (July 2026)

Anthropic prices the API per million tokens for both input and output. Output tokens always cost more because they require more compute. Here are the current rates for every generally available model.

Sonnet 5 has introductory pricing through August 31, 2026. After that date it reverts to the standard $3/$15 rate. Haiku 4.5 is the most affordable option for high-volume, latency-tolerant tasks.

Prompt Caching Cuts Input Costs by 90%

Prompt caching lets you reuse large system prompts, documents, or tool definitions across requests without paying full price each time. Anthropic applies two separate rates to cached content.

Cache writes — the first time a prompt block enters the cache — cost 25% more than the base input rate. Every subsequent cache read of that same block gets a 90% discount off the base input rate. For a system prompt you send thousands of times, the savings compound fast.

In practice, most teams see a 70-85% reduction in input token costs after enabling caching on their largest prompt components. The cache has a five-minute TTL that resets on each hit, so active workloads keep their cache warm automatically.

• Cache write: base input price + 25%
• Cache read: base input price - 90%
• Five-minute TTL, refreshed on each cache hit
• Works on system prompts, tool schemas, and any reusable prefix
• Available on all Claude models via the API

Batch API: 50% Off for Async Workloads

The Batch API offers a flat 50% discount on both input and output tokens. The tradeoff is turnaround time — results arrive within 24 hours instead of streaming in real time.

This makes batching ideal for offline jobs like document classification, data extraction, content generation, and evaluation pipelines. You submit a batch of requests as a single JSONL file and poll for results.

Batch pricing stacks with prompt caching. If you cache your system prompt and run a batch job, you get both the 90% cache-read discount on input and the 50% batch discount on output.

• 50% discount on all token costs (input and output)
• Results delivered within 24 hours
• Submit via JSONL file with up to 10,000 requests per batch
• Stackable with prompt caching for maximum savings
• Best for evaluation, extraction, and bulk content tasks

Extended Thinking Token Costs

Extended thinking lets Claude reason through complex problems step by step before giving its final answer. The thinking tokens Claude generates are billed at the output token rate for the model you are using.

For Opus 5 at $25 per million output tokens, a response that uses 2,000 thinking tokens adds about $0.05 to your cost. Most extended thinking responses use between 500 and 5,000 thinking tokens depending on problem complexity.

You can set a thinking budget to cap costs. The budget parameter limits how many thinking tokens Claude will use before it must produce a final answer. This gives you direct control over the cost-quality tradeoff.

• Thinking tokens billed at the model's output rate
• Controllable via a thinking budget parameter
• Typical range: 500-5,000 thinking tokens per response
• Most valuable for math, code, and multi-step reasoning tasks

Claude Code Plans vs. API Pay-Per-Token

Anthropic offers two ways to pay: flat-rate Claude Code plans or direct API billing. Each suits a different use case.

The Max plan costs $100 per month with Opus 4.7 as the default model, or $200 per month with Opus 5. You get unlimited usage within fair-use limits, which works well for individual developers who use Claude Code daily. The Team plan costs $30 per seat per month with shared billing and admin controls. Enterprise pricing is custom.

The API is better when you are building applications that call Claude programmatically. You pay only for the tokens you consume, with no per-seat charge. For bursty workloads — heavy one week, quiet the next — API billing is usually cheaper than a flat plan.

• Max plan: $100/mo (Opus 4.7) or $200/mo (Opus 5) — flat rate
• Team plan: $30/seat/month with shared admin and billing
• Enterprise: custom pricing with SSO, audit logs, and SLAs
• API: pure pay-per-token, no minimum, no seat charge
• Choose plans for daily individual use, API for programmatic workloads

How to Estimate Your Monthly Claude API Spend

Start by measuring two numbers: your average input tokens per request and your average output tokens per request. Multiply each by your expected monthly request volume, then apply the per-million rate for your chosen model.

A typical customer-support chatbot sending 500-token prompts and receiving 300-token responses on Haiku 4.5 costs roughly $0.65 per 1,000 conversations. A code-generation pipeline using Opus 5 with 2,000-token prompts and 1,500-token outputs runs about $4.75 per 100 requests.

Add prompt caching to the estimate. If 80% of your input tokens are cacheable and you enable caching, replace those tokens with the cache-read rate (90% off). This single optimization often cuts the total bill in half.

• Formula: (input tokens x input rate) + (output tokens x output rate) = cost per request
• Multiply cost per request by monthly volume for total spend
• Factor in caching: replace cacheable input tokens with 90%-discounted rate
• Use the Batch API estimate for any async workloads (50% off)
• Anthropic's usage dashboard tracks spend in real time

Seven Ways to Reduce Claude API Costs

Lowering your API bill does not mean sacrificing quality. These strategies target waste and take advantage of built-in discounts.

• Enable prompt caching on system prompts, tool definitions, and any reusable context — this alone can save 70-85% on input costs
• Use the Batch API for any workload that does not need real-time responses — saves 50%
• Pick the cheapest model that meets your quality bar — Haiku 4.5 handles classification and extraction well at $1/$5
• Trim input tokens by removing unnecessary whitespace, redundant instructions, and conversation history older than the relevant window
• Set a thinking budget on extended thinking to prevent runaway reasoning costs
• Route requests by complexity: use Haiku for simple tasks and Opus only when the task demands it
• Monitor usage via the API dashboard and set spend alerts to catch unexpected spikes

Bedrock and Vertex AI Pricing Differences

Claude is also available through Amazon Bedrock and Google Cloud Vertex AI. Both platforms set their own pricing, which may differ from Anthropic's direct API rates.

Bedrock and Vertex AI pricing typically includes a small markup over direct API rates. However, they offer benefits like consolidated cloud billing, reserved capacity, and integration with existing cloud infrastructure. If your team already runs on AWS or GCP, the operational simplicity may justify the premium.

Check the current Bedrock and Vertex AI pricing pages for exact rates, as they update independently from Anthropic's direct pricing.