Claude API Pricing Breakdown (2026): Tokens, Limits & Costs

Executive Summary

The Claude Developer Platform from Anthropic prices the API primarily by input tokens and output tokens, with cost modifiers for prompt caching, batch processing, long-context requests, and data residency routing. The primary endpoint is Messages (POST /v1/messages) on api.anthropic.com.

For the latest trio (Opus/Sonnet/Haiku), official list prices (standard speed, ≤200K input) are: Opus 4.6 $5/MTok in, $25/MTok out; Sonnet 4.6 $3/MTok in, $15/MTok out; Haiku 4.5 $1/MTok in, $5/MTok out. Output tokens are 5× the input rate, so response length is often the biggest cost lever.

For the chatbot workload (100k messages/month, 1k tokens/message, 75% input / 25% output), estimated monthly spend is ~$1,000 (Opus), ~$600 (Sonnet), or ~$200 (Haiku) at standard pricing.

Two gotchas: Embeddings and fine-tuning are not first-party — Anthropic does not offer its own embedding model, and the API does not currently offer fine-tuning. Tooling adds non-token billing: web search is $10/1,000 searches on top of token costs.

How Claude API Pricing Works

Claude bills for input and output tokens. Cost modifiers include:

• Prompt caching: Cache writes at 1.25× or 2× base; cache reads at 0.1× base
• Batch API: 50% discount on input and output
• Long context: Requests >200K input tokens use premium long-context rates
• Data residency: US-only inference (inference_geo) incurs 1.1× multiplier for Opus 4.6 and newer

Token Pricing Table

Caching pays off quickly: 1 cache read for 5-minute cache, 2 reads for 1-hour cache. Multipliers stack with batch discounts and data residency.

Batch API: 50% Cost Reduction

Anthropic offers a 50% discount for the Batch API on both input and output tokens:

Batch is ideal for offline workloads (evaluations, bulk summarisation, classification). Batches are limited to 100,000 requests or 256MB; results available for 29 days. Fast mode is not available with Batch.

Model Families and Capabilities

• Opus: Highest capability for complex agentic tasks and coding. 1M context (beta), up to 128K output.
• Sonnet: Balance of speed and intelligence. Often the best default for most applications.
• Haiku: Fastest and most economical for high-volume, cost-sensitive deployments.

Extended thinking (Opus 4.6, Sonnet 4.6): thinking budget tokens are billed as output. Enabling extended thinking can materially increase output cost.

Vision: Image Token Estimation

Claude supports image + text input. Image tokens are estimated as:

image_tokens ≈ (width_px × height_px) / 750

A 1000×1000 image ≈ 1,334 tokens. Resize images below thresholds to avoid unnecessary latency and cost.

Real Cost Examples

Chatbot: 100k messages/month, 1k token/msg, 75% input / 25% output

75M input + 25M output tokens monthly.

Multimodal: 10k images/month (1000×1000, 50 text in + 200 out)

~1,334 image tokens + 50 text = 1,384 input per request.

Embeddings

Anthropic does not offer a first-party embedding model. Use an external provider (e.g. Voyage AI) for RAG and retrieval pipelines. Embedding costs are not part of the Claude bill.

Sensitivity: Token Volume and Output Share

At 100k messages/month, cost varies with tokens per message and output share:

Shifting output from 15% to 40% increases cost significantly.

Hidden Costs and Tooling

Web search: $10 per 1,000 searches, plus standard token costs for search-generated content.

Tool overhead: Enabling tools adds system prompt tokens (~313–346 depending on config). Tool schemas and tool result blocks count as input/output.

Code execution: Free when used with web search or web fetch; otherwise billed by execution time with a minimum runtime and monthly free allowance.

Long-context premium: If you enable 1M context and total input exceeds 200K tokens, the entire request switches to premium long-context pricing. The 200K threshold includes cache reads/writes.

Data residency: US-only inference adds 1.1× multiplier for Opus 4.6 and newer models.

Rate Limits

Anthropic measures limits in RPM (requests/minute), ITPM (input tokens/minute), and OTPM (output tokens/minute). Cached read tokens typically do not count toward ITPM. Exact values vary by tier and model — check the console for your project.

Tier progression: Tier 1 $5 cumulative credit; Tier 2 $40; Tier 3 $200; Tier 4 $400.

How to Reduce Claude API Costs

• Model routing: Default to Haiku or Sonnet; escalate to Opus only for complex reasoning, high-stakes coding, or hard tool orchestration.
• Output control: Output is 5× input price. Shorter answers, structured outputs, and retrieval-first responses yield outsized savings.
• Prompt caching: Cache large stable instructions, tool catalogs, and long documents. Cached reads are 0.1× base and often don't count toward ITPM.
• Batch API: Use for non-interactive workloads — 50% cost reduction, most batches complete in under an hour.
• Image resizing: Keep images below resizing thresholds; use the token formula to budget vision costs early.

Related

For a side-by-side comparison with OpenAI and Gemini, see AI API Pricing Comparison (2026).