🟢 Model references match the current lineup · model notice · Fable subscription
Fable 5 and 5.1 subscription (updated September 7, 2026): Claude Fable 5.1, released September 1, 2026, is the current Fable model and Fable 5 is now legacy. Plan terms are the same for both — Max and Team Premium plans include Fable at up to 50% of the weekly usage limit; Pro and Team Standard use usage credits (
Update (September 27, 2026): Claude Opus 5.5 was released on September 22, 2026 and is now the current Opus model (API claude-opus-5-5, $4 per million input tokens and $20 per million output tokens, down from Opus 5's $5/$25). Anthropic says it performs at the level of Fable 5.1 on most work and costs 40% less to run than Opus 5 on typical workloads. Opus 5 remains available but is marked legacy in the official model docs. Opus 5 prices and comparisons below reflect the pre-launch state and will be updated. Details: Claude Opus 5.5 complete guide.
This article covers what you pay per token when calling the Claude API (a programming interface for connecting Claude to your own app or service) — not the same thing as Claude.ai's Free, Pro, or Max subscription fees, which follow a completely different pricing structure. We'll compare base rates across models, then show how the Batch API's 50% discount and prompt caching's up to 90% discount work, with worked examples. About a 5-minute read.
What does the Claude API charge per model?
As of September 2026, base Claude API pricing ranges from $1 to $10 per million input tokens and $5 to $50 per million output tokens, depending on the model tier. Pricing is quoted per MTok (million tokens), the standard billing unit. Here are the current per-model rates.
| Model | Input (per MTok) | Output (per MTok) | Context window | Best for |
|---|---|---|---|---|
| Claude Haiku 4.5 | $1 | $5 | 200K | Fastest and cheapest, simple repetitive tasks |
| Claude Sonnet 5 | $2 | $10 | 1M | Best balance of speed and intelligence, most production workloads |
| Claude Opus 5 | $5 | $25 | 1M | Complex agentic coding and enterprise work |
| Claude Fable 5.1 | $10 | $50 | 1M | Anthropic's most capable model, per official docs |
The context window — the maximum amount of input a model can consider at once — is 1 million tokens for Opus 5, Sonnet 5, and Fable 5.1, and 200,000 tokens for Haiku 4.5.
Why do input and output prices differ across models?
Output tokens cost more than input tokens because generating new text is computationally heavier than reading text that's already provided. As the table above shows, output pricing is set at exactly 5x input pricing across all four models — so response length affects your bill much faster than prompt length does.
One caveat worth noting: per official docs, Opus 4.7 and later Opus models, Fable 5, and Sonnet 5 use a newer tokenizer (the method used to split text into tokens) that produces roughly 30% more tokens for the same text compared to earlier-generation models. So comparing per-token rates alone isn't enough — the actual token count for the same content can differ by model generation too.
How does the Batch API cut costs in half?
The Batch API — an asynchronous mode for processing large volumes of requests that don't need an immediate response — applies a 50% discount to both input and output token pricing. It's a good fit for bulk document summarization, data labeling, or overnight batch analysis where waiting minutes to hours for results is fine.
| Model | Batch input (per MTok) | Batch output (per MTok) |
|---|---|---|
| Claude Haiku 4.5 | $0.50 | $2.50 |
| Claude Sonnet 5 | $1 | $5 |
| Claude Opus 4.8 | $2.50 | $12.50 |
| Claude Fable 5.1 | $5 | $25 |
How does prompt caching reduce the cost of repeated context?
Prompt caching lets you store a long system prompt or document you send repeatedly in a cache — a temporary storage layer for reusing frequently-used data quickly — so later requests pay only 10% of the base input price to reuse it. Storing content in the cache for the first time (a cache write) actually costs more than a normal input token: 1.25x for a 5-minute cache, or 2x for a 1-hour cache.
In short: a 5-minute cache pays for itself after just one reuse (write 1.25x + one read 0.1x = 1.35x, cheaper than sending the same content twice at 2x), while a 1-hour cache needs at least two reuses to break even. There are two ways to turn caching on: automatic caching, where a single cache_control field at the top of your request lets the system manage cache boundaries for you, and explicit cache breakpoints, where you tag specific content blocks yourself for finer control. If you're just starting out, automatic caching is the simpler option to try first.
Can you combine Batch API and prompt caching?
Yes, Batch API and prompt caching discounts can be combined. Official docs note that using both together provides significant savings compared to standard API calls. One more thing worth knowing: if you pin data processing to a specific region (data residency), a separate 1.1x multiplier applies across input, output, and cache pricing.
What does this look like in real costs?
Official documentation offers a concrete example: processing 10,000 customer support tickets averaging 3,700 tokens per conversation with Claude Haiku 4.5 ($1 input / $5 output) costs roughly $37 total.
You can work out caching savings yourself too. Say you repeatedly send a 100,000-token document as a system prompt with Sonnet 5 ($2/MTok input). The first cache write (5-minute TTL) costs 100,000 × $2.5 ÷ 1,000,000 = $0.25. Every subsequent reuse of that cache costs 100,000 × $0.2 ÷ 1,000,000 = $0.02 — a tenth of the $0.20 it would cost to resend the same content as a normal input each time.
Which model should you use, and when?
Official guidance recommends Haiku for simple tasks, Sonnet for most production workloads, and Opus for complex agentic coding and enterprise work, with Fable 5.1 positioned as the choice when a workload needs the highest available capability for long-running agents. For cost-sensitive, high-volume workloads, layering in Batch API and prompt caching on top of the right model is the recommended way to bring your actual bill down.
Frequently Asked Questions
Q. Can you use prompt caching and Batch API at the same time?
Yes, official documentation confirms the two discounts stack. Combining them gives the biggest savings on large-volume, repeated-context workloads.
Q. How long does a cache last?
You can choose either 5 minutes or 1 hour. Once the TTL (Time To Live — how long a cache stays valid) expires, the cache is gone, and the next request pays the cache-write price again.
Q. Is Sonnet 5 pricing about to change?
No. $2 input / $10 output was announced as introductory pricing through August 31, 2026, with $3 / $15 to follow from September 1 — but Anthropic cancelled that increase and confirmed $2 / $10 as the standard price (official pricing docs, checked September 2, 2026).
Q. Is Claude.ai's subscription fee the same as API pricing?
No, they're separate systems. Free, Pro ($20/month), and Max ($100 or $200/month) are app subscriptions with fixed usage limits, while the API pricing covered here is a separate pay-per-token system for developers.
Combining per-model rates with Batch API and prompt caching can meaningfully cut the cost of the same workload. Before applying any of this to production, it's worth double-checking current figures on the official pricing pages (claude.com/pricing, platform.claude.com).
Curious what your usage would cost? Try the API cost calculator — pick a model and token counts for an instant estimate.
Related articles
- Claude API Model IDs and Versioning: Aliases, Pinned Snapshots, and Migration
- Claude API Context Windows and Tokens: Limits and Management
- Claude Model Comparison: Opus 5, Sonnet 5, Haiku 4.5, Fable 5.1
- Claude Pricing 2026: Free vs Pro vs Max — Which to Pick?
- Getting Started with the Claude API: Your First Call (Developer Intro)