Claude API Pricing 2026: Per-Token Rates, Real Cost Examples, and Free API Credits

All prices in USD per million tokens, from Anthropic's pricing docs. Cache write columns are what you pay to store a prefix; cache read is what you pay when a later request reuses it.

The Batch API processes requests asynchronously, typically within hours, at half price on both input and output. It stacks with prompt caching. Fast mode is not available in batch.

List price is the starting point. These are the modifiers Anthropic applies, and most production bills are shaped more by them than by the model choice.

Prompt caching

A 5-minute cache write costs 1.25x the input price; a 1-hour write costs 2x. Every read of that cached prefix costs 0.1x input (0.025x on Fable 5.1). So caching pays for itself after one read on the 5-minute TTL and after two reads on the 1-hour TTL. For any app with a long system prompt or shared context, caching is the difference between a viable margin and not.

Long context

Claude 4.6 and later include the full 1M-token context window at standard pricing. A 900k-token request is billed at the same per-token rate as a 9k-token request. There is no long-context surcharge.

Tokenizer change

Claude 4.7 and later use a newer tokenizer that produces roughly 30% more tokens for the same text. Sonnet 5 at $2/$10 is still cheaper than Sonnet 4.6 at $3/$15 after that, but compare on your real prompts rather than on list price.

Fast mode

Fast mode on Opus 5 and Opus 4.8 gives significantly faster output at $10 input / $50 output, double the standard Opus rate. It stacks with caching and data residency multipliers and is first-party API only.

Data residency

US-only inference (inference_geo: "us") on 4.6+ models adds a 1.1x multiplier to every token category. Global routing, the default, uses standard pricing. On Bedrock and Google Cloud, regional endpoints carry a 10% premium over global ones.

Tool use and server-side tools

Tool definitions count as input tokens, and each model adds a tool-use system prompt of a few hundred tokens. Web search is $10 per 1,000 searches plus tokens. Web fetch is free beyond tokens. Code execution is free when used with web search or web fetch; otherwise each org gets 1,550 free container-hours a month, then $0.05 per hour.

Managed Agents

Claude Managed Agents bill tokens at the rates above plus $0.08 per session-hour of runtime while the session is in the running state. Batch discounts do not apply to agent sessions.

Real arithmetic using the rates above. Token counts are typical for each workload; substitute your own.

Rate limits are set by usage tier, not by plan. Anthropic describes three standard tiers: Start (entry-level limits), Build (increased limits for growing apps) and Scale (the highest standard limits for production). You move up by spending and by time on the platform, and you can request limits beyond Scale through sales.

Volume discounts exist but are negotiated case by case, and Anthropic notes that academic and research discounts may be available. Enterprise customers get custom rate limits, dedicated support and invoicing. If you are spending five figures a month, ask.

Billing is monthly in arrears on actual usage, by credit card or invoice, and every dollar is visible in the Claude Console. If you buy through AWS Marketplace or Azure Marketplace, usage is metered in Claude Consumption Units at $0.01 each, converted from the same per-token rates after any negotiated discount.

Claude for Startups members who receive credits automatically get higher API rate limits, which is often worth as much as the credits themselves when you are launching.

Order matters. Set up the Console before you apply so the credits have somewhere to land.