How Much Does Anthropic API Cost in 2026? Full Pricing Breakdown
The Anthropic API has no subscription fee — you pay per token, and in 2026 current model rates run from $0.10 input / $0.50 output per million tokens (Claude Haiku 5.5) up to $10/$50 for Claude Fable 5.1, with the workhorse Sonnet 5.5 at $2/$10 and flagship Opus 5.5 at $4/$20. Prompt caching cuts repeated input by up to 97.5% (reads at 0.025x–0.1x base input) and the Batch API takes another 50% off, so cache-heavy, asynchronous workloads can cost a fraction of sticker price. If you're building agent workflows on Claude, budget by model tier first — model choice spans a 100x input-price range within Anthropic's own lineup.
TL;DR
- The Anthropic API itself has no subscription fee — billing is per token, split into input and output, quoted per million tokens (MTok).
- Current 2026 rates: Claude Haiku 5.5 $0.10/$0.50, Sonnet 5.5 $2/$10, Sonnet 5 $2/$10, Opus 5.5 $4/$20, Opus 5 $5/$25, Fable 5.1 $10/$50 per 1M input/output tokens.
- Output tokens cost 5x input at Anthropic, so verbose responses dominate most bills — model choice spans a 100x input-price range.
- Prompt caching: 5-minute cache writes cost 1.25x base input, 1-hour writes 2x; cache reads run 0.025x–0.1x base input depending on model.
- The Batch API halves input and output for asynchronous jobs and stacks with caching — a batched Sonnet 5.5 cache hit costs $0.05/M.
- Claude subscription plans (Pro $17/mo annual, Max from $100/mo) are separate from the API; Max 5x and 20x include $100–$200/month in API credits.
- Anthropic made Sonnet 5's introductory $2/$10 rate permanent in August 2026 rather than letting it rise to $3/$15.
Head to head
| Dimension | Anthropic API |
|---|---|
| Pricing model | Per-token usage; no API subscription fee 🏆 |
| Entry input rate | $0.10/MTok (Claude Haiku 5.5) 🏆 |
| Flagship input rate | $4–$5/MTok (Opus 5.5 / Opus 5); Fable 5.1 at $10 🏆 |
| Caching discount | Reads at 0.025x–0.1x base input; writes 1.25x (5-min) or 2x (1-hour) 🏆 |
| Batch discount | 50% off input and output; stacks with caching 🏆 |
| Context window | 1M tokens at standard pricing on Sonnet/Opus tiers; Haiku 4.5 caps at 200K 🏆 |
| Cost transparency | Token rates public to the cent; subscription usage caps unstated in tokens 🏆 |
| Commitment | None for the API; Max plans bundle $100–$200/month API credits 🏆 |
The contenders
Anthropic's API provides programmatic access to the Claude model family and is billed purely on usage: you pay per token, split into input (what you send) and output (what the model writes back), quoted per million tokens. As of late 2026 the current lineup spans Claude Haiku 5.5 ($0.10 input / $0.50 output per MTok), Sonnet 5.5 and Sonnet 5 ($2/$10), Opus 5.5 ($4/$20), Opus 5 ($5/$25), and the premium Fable 5.1 ($10/$50), with 1M-token context windows on the mid and upper tiers. Three discount levers materially change real-world spend: prompt caching (5-minute writes at 1.25x base input, 1-hour at 2x, reads as low as 0.025x), the Batch API (50% off input and output for asynchronous jobs, up to 100,000 requests or 256 MB per batch), and model choice itself — the spread from Haiku 5.5 to Fable 5.1 is 100x on input. Anthropic's consumer/Team plans (Free, Pro at $17/month billed annually, Max from $100, Team at $20/seat/month annually) are separate products that do not pool into API tokens, though Max plans include $100–$200/month in API credits. For small businesses, the API's appeal is predictable, published rates with no minimum commitment; the trade-off is that output tokens cost 5x input, so uncontrolled verbosity is the biggest budget risk.
pricing 0.1/month
- + Rates published to the cent for every model, with batch and cache discounts stacking for up to ~95%+ effective savings
- + No subscription fee or minimum spend on the API itself — viable for very small workloads
- + Wide model ladder (100x input-price spread) lets small businesses match tier to task
- + Sonnet 5's $2/$10 rate was made permanent in August 2026 rather than increasing to $3/$15
- + 1M context at no long-context surcharge on Sonnet/Opus tiers, unlike some competitors
- − Output tokens cost 5x input, so verbose or agentic workloads escalate quickly
- − Cache writes cost extra (1.25x–2x base input) and caches have minimum prompt lengths; a marked prefix below threshold processes uncached without error
- − Pro/Max subscription usage caps are expressed as multiples of an unstated token allowance, making subscription-vs-API comparisons hard to verify
- − New model releases can change token consumption — Opus 4.7's tokenizer could use up to 35% more tokens for the same text
- − Fast Mode does not work with the Batch API, limiting discount stacking on latency-sensitive premium work
Verdict
Winner: Anthropic API
As the only entity on this page, the Anthropic API's 2026 pricing is evaluated on its own terms: transparent per-token rates, steep caching and batch discounts, and a wide model ladder from $0.10 to $10 per million input tokens.
| Option | Overall | Breakdown |
|---|---|---|
| Anthropic API | 8.5 | : 9: 9: 9: 7: 8: 8 |
Anthropic API pricing in 2026: the full rate card
Caching, batching, and the discounts that change everything
API versus Claude subscription plans
Which should you pick?
- :
- :
- :
- :
- :
FAQ
Does the Anthropic API charge a subscription fee?
No. The API key itself has no subscription fee; your bill is determined by the selected model, token counts, tools used, and pricing modifiers such as caching and batch processing. Anthropic's subscription products (Free, Pro at $17/month annual, Max from $100, Team at $20/seat/month annual) are separate and do not pool into standard API tokens, though Max plans include $100–$200/month in API credits.
What is the cheapest Claude model on the API in 2026?
Claude Haiku 5.5, at $0.10 per million input tokens and $0.50 per million output tokens, with a 1M-token context window and batch rates of $0.05/$0.25. It's a dramatic drop from the previous budget tier, Haiku 4.5 at $1/$5, and sits far below the flagship Opus 5.5 at $4/$20 and premium Fable 5.1 at $10/$50.
How much do prompt caching and batch processing actually save?
Cache reads cost between 0.025x and 0.1x the base input rate depending on model (Fable 5.1 $0.25/M, Opus 5.5 $0.20/M, Sonnet 5.5 $0.10/M), though writes cost 1.25x (5-minute) or 2x (1-hour) base input. The Batch API takes 50% off input and output and stacks with caching — a batched Sonnet 5.5 cache hit costs $0.05/M. Combined, effective savings can exceed 95% on cache-heavy asynchronous workloads.
Why is my Anthropic bill higher than the sticker rate suggests?
Three common causes: output tokens cost 5x input, so verbose responses dominate; cache writes are billed at 1.25x–2x base input before any read savings; and tool-use system prompts add fixed token overhead (for example, 497 tokens on Sonnet 4.6 per official docs). Also capture every billable call in an agent workflow — measuring only the final response misses earlier model calls.
Did Anthropic API prices change in 2026?
Headline rates were remarkably stable — Opus-tier models held $5/$25 across the 4.6, 4.7, 4.8 and 5 generations — but two shifts mattered. Anthropic made Sonnet 5's introductory $2/$10 rate permanent on August 11, 2026, cancelling a planned increase to $3/$15. And new releases can change consumption: Opus 4.7's new tokenizer could use up to 35% more tokens for the same text, even at the same per-token price.