Anthropic API Pricing in 2026: Complete Guide- Models, Caching, Batch & Optimization

Already using Claude at scale? See how Finout tracks it

Model Profiles: Capabilities & When to Use Each

Claude Fable 5.1: Most Capable

API ID: claude-fable-5-1 | Price: $10.00/$50.00 per MTok | Context: 1M tokens

Anthropic positions Fable 5.1 for demanding reasoning and long-horizon agentic work, and recommends it when evals on Opus 5.5 at higher effort still fall short. Its cache hit rate is $0.25 per MTok, just 2.5% of base input. That puts cached input within 1.25× of Opus 5.5 ($0.20), even though uncached input is 2.5× the price. On cache-heavy agent workloads, the real gap between the two models comes mostly from output tokens.

Best for: The hardest reasoning and long-running agentic tasks, where Opus 5.5 has already been tested and falls short.

Claude Opus 5.5: Recommended Starting Point

API ID: claude-opus-5-5 | Price: $4.00/$20.00 per MTok | Context: 1M tokens

Anthropic’s model overview tells developers who are unsure which model to use to start with Opus 5.5 for most workloads. It costs 20% less than Opus 5 per token, and its cache hits cost $0.20 per MTok compared to $0.50 on Opus 5.

Fast Mode on Opus 5.5 costs $8/$40 per MTok, twice the standard rate.

Migration from Opus 5 or Opus 4.8: Both already use the newer tokenizer, so moving to Opus 5.5 lowers cost per request directly. Migration from Opus 4.6 or earlier: the roughly 30% increase in token count can offset the 20% price cut. In a workload that sees the full increase, cost per request comes out about 4% higher. Benchmark on your own traffic before you switch.

Best for: Long-running agentic coding and knowledge work, and the default choice when you have not yet benchmarked other tiers.

Claude Sonnet 5.5 and Sonnet 5

API IDs: claude-sonnet-5-5, claude-sonnet-5 | Price: $2.00/$10.00 per MTok | Context: 1M tokens

Both Sonnet models cost half as much as Opus 5.5 per token. Sonnet 5.5 cuts the cache hit rate in half compared to Sonnet 5 ($0.10 vs $0.20 per MTok). For cache-heavy workloads, it is the cheaper of the two at the same list price.

Migration from Sonnet 4.6: Per-token prices drop by a third ($2/$10 vs $3/$15), but Sonnet 4.6 uses the older tokenizer. If your workload sees the full ~30% increase in token count, cost per request still comes out roughly 13% lower.

Best for: The majority of production workloads where Opus-level reasoning is not required.

Claude Haiku 5.5 and Haiku 4.5: Budget

API IDs: claude-haiku-5-5, claude-haiku-4-5-20251001 | Price: $0.10/$0.50 (Haiku 5.5, prompts up to 100K) and $1.00/$5.00 (Haiku 4.5)

Haiku 5.5 is the cheapest model in the lineup by a wide margin, but its pricing depends on prompt length. Any request whose prompt goes over 100,000 tokens pays 5× the per-token rate for the whole request, including cache reads. Keep prompts under the threshold to get the low rate.

Best for: High-volume, short-prompt workloads such as classification, routing, extraction, summarization, and moderation.

Prompt Caching: The Biggest Cost Lever Available

Prompt caching is Anthropic's most impactful pricing feature. It lets you store frequently reused content, such as system prompts, documents, examples, tool definitions, so subsequent API calls can read from cache instead of reprocessing the full input. Cache hits cost 90% less than standard input on most models, 95% less on Opus 5.5 and Sonnet 5.5, and 97.5% less on Fable 5.1.

How Prompt Caching Works

There are two ways to enable it. Automatic caching adds a single cache_control field at the top level of the request and lets the system manage cache breakpoints as conversations grow. Anthropic recommends this as the starting point for most use cases. Explicit breakpoints place cache_control on individual content blocks for fine-grained control. On the first request, Anthropic processes and stores the cached content. The write costs 1.25× standard input for a 5-minute TTL, or 2× for a 1-hour TTL. Subsequent requests within the TTL window pay the cache hit rate. Reading a cached block refreshes it, so high-frequency applications rarely pay write costs after the initial warm-up. At the standard 0.1× hit rate, a 5-minute cache pays for itself after one read and a 1-hour cache after two.

Prompt Caching Pricing by Model

Caching Cost Example: RAG Application

Scenario: A RAG app with a 50K-token knowledge base in the system prompt, queried 1,000 times per day on Sonnet 5.5. For simplicity, this example counts only the knowledge base tokens. User questions and output cost the same with or without caching.

Without caching: 1,000 queries/day at 50K tokens each

• Input tokens per day: 50M tokens
• Standard input cost ($2.00/MTok): $100.00/day
• Monthly cost: ~$3,000/mo

With caching: same workload, 1-hour cache rewritten once an hour

• 24 cache writes per day ($4.00/MTok × 50K tokens × 24): $4.80/day
• 976 cache hits per day ($0.10/MTok × 50K tokens × 976): $4.88/day
• Monthly cost: ~$290/mo
• Daily spend: ~$9.68 (about 90% lower)

Key insight: Prompt caching is not just an optimization. For any application with a large, reused system prompt or document context, it is the single most impactful change you can make to your Anthropic bill. Applications that query the same knowledge base repeatedly should always use caching.

Batch Processing: 50% Off for Async Workloads

The Anthropic Message Batches API processes requests asynchronously and returns results within 24 hours, at exactly 50% off standard token prices. There is no quality difference between batch and real-time responses, only timing. Batch and caching discounts stack.

Best workloads for batch processing: document processing pipelines, data enrichment at scale, nightly analytics jobs, offline evaluations, content generation queues, and any task where a few hours of latency is acceptable. 

Long-Context Pricing: What Changes With Prompt Length

Claude 4.6 and later models include the full 1M token context window at standard pricing, so a 900K-token request is billed at the same per-token rate as a 9K-token request. The one exception in the current lineup is Haiku 5.5.