Anthropic API Pricing: Claude Model Costs and the Cheapest Sensible Route
Anthropic bills API and Workbench usage through prepaid credits. Claude model pricing spans premium Opus, balanced Sonnet and low-cost Haiku tiers. The largest savings usually come from model routing and prompt caching rather than choosing one model for every request.
By Aniruddh PanvalkarPricing verified Oct 18, 2026
Plans and pricing
Claude Opus tier
From US$15 / 1M input and US$75 / 1M output on the published Opus rate cardPremium Claude tier for the hardest reasoning and high-value agentic work.
Best for
Tasks where accuracy, planning depth or failure avoidance justifies premium output cost.
Not for
Routine production traffic, simple generation or cost-sensitive batch work.
Claude Sonnet tier
From US$3 / 1M input and US$15 / 1M output on the published Sonnet rate cardBalanced Claude tier widely used for coding, writing and production agents.
Best for
Coding, document work, structured outputs and general production workloads.
Not for
Very simple high-volume tasks that Haiku can pass.
Claude Haiku tier
Published Haiku rates vary by generation; verify the current model tableLower-cost Claude tier for latency-sensitive and high-volume workloads.
Best for
Classification, extraction, routing, short answers and bulk automation.
Not for
Hard reasoning or complex multi-file coding where Sonnet or Opus materially improves success.
Anthropic API credits are prepaid, expire after one year and are non-refundable. Model rates and available generations change; verify Anthropic's official pricing before purchase.
Which Claude API tier should you use?
Use Haiku for cheap routine work, Sonnet as the production default when quality matters, and Opus only for tasks that demonstrate a measurable success-rate improvement. Test caching on repeated context before assuming raw token rates equal landed cost.
Calculate your exact cost →Who is overpaying for Anthropic API?
- Teams using Opus for work that Sonnet passes reliably
- Agents repeatedly sending large unchanged prompts without caching
- Organizations prepaying more credits than they can use before expiry
Sources
Prices on this page are checked against these vendor pricing pages and catalogue entries.
- anthropic.com/pricing
- docs.anthropic.com/en/docs/about-claude/pricing
- support.anthropic.com/en/articles/8977456-how-do-i-pay-for-my-api-usage
- docs.anthropic.com/en/docs/about-claude/models
- claude.ai
- anthropic.com/claude-code
- openrouter.ai/anthropic/claude-fable-5
- openrouter.ai/anthropic/claude-fable-5:batch
- openrouter.ai/anthropic/claude-fable-5.1
- openrouter.ai/anthropic/claude-fable-5.1:batch
- openrouter.ai/anthropic/claude-haiku-4.5:batch
- openrouter.ai/anthropic/claude-haiku-5.5
Frequently asked questions
How does Anthropic API billing work?
Self-serve API and Workbench usage is deducted from prepaid usage credits. Failed requests are generally not charged, and credits expire one year after purchase.
What is the main Claude API cost-saving tactic?
Route simple tasks to Haiku, keep Sonnet for production-quality work, reserve Opus for proven hard cases, and use prompt caching for repeated context.