Claude API Pricing Comparison 2026

Every Claude request is billed by tokens, not by message or by minute. A token is a small piece of text, usually part of a word. Four kinds of tokens show up on a bill:

• Input tokens: everything you send: the system prompt, the conversation history, tool definitions, documents and images.
• Output tokens: everything Claude writes back, including tool calls and any visible reasoning.
• Cache write tokens: input that you mark for prompt caching the first time it is stored. These cost a little more than normal input.
• Cache read tokens: cached input reused by a later request. These cost a small fraction of the normal input price.

Output is the expensive side. For every Claude model, output costs about five times as much as input: on Claude Sonnet 4.6 that is $5.74 versus $1.15 per million on APIVAI. So a request with a long prompt and a short answer is cheaper than its size suggests, and a request that generates long code files is more expensive than it looks.

On APIVAI you pay from a prepaid USD balance, per request, for exactly these token types. There is no subscription and no monthly minimum, and the dashboard lists every request with its tokens and cost.

Prices as of 2026-10-10. Official prices are Anthropic's list prices from anthropic.com/pricing; APIVAI prices update automatically on the pricing page.

APIVAI cache prices per million tokens for the main models, as of 2026-10-10. For Anthropic's official cache prices, see anthropic.com/pricing.

Rough rules that are good enough for a budget:

• English text averages about 4 characters per token, or roughly three quarters of a word per token. A 1,000-word document is around 1,300 tokens.
• Code usually uses more tokens per line than prose, because of symbols, indentation and short identifiers.
• Chinese, Japanese and other non-Latin scripts generally use more tokens per character than English. Treat estimates for these languages as approximate and measure real samples.
• Each turn of a chat resends the whole history, so input grows with conversation length.

For an exact count before you send anything, call the token counting endpoint. It returns the number of input tokens a request would use without generating a response:

The response contains input_tokens. Run it on a few typical requests, multiply by your monthly volume, then check the totals against the dashboard after the first days of real traffic. The docs list all endpoints.

A help-desk bot handles 10,000 conversations a month. Each conversation averages 3,000 input tokens (system prompt, a few knowledge-base snippets and the chat history) and 400 output tokens.

Haiku handles routine questions well; many teams send only escalated or complex tickets to Sonnet. Caching the system prompt lowers these numbers further, because it is identical in every conversation.

A developer runs 20 Claude Code sessions a day, each about 50,000 input and 20,000 output tokens, 30 days a month (600 sessions). Setup takes two environment variables; see the Claude Code setup guide.

This ignores caching. In practice Claude Code caches most of the repeated context (files, instructions, earlier turns), which lowers both columns. For a comparison with subscription-based editors, read Claude Code vs Cursor: cost comparison.

A one-off job summarizes 1,000 reports of about 20,000 tokens each into a 1,000-token summary.

Here input dominates, so the per-input-token price matters most. APIVAI runs these as normal requests; it does not offer a Batch API. Anthropic's official Batch API has its own discounted pricing for jobs that can wait, so compare both on Anthropic's pricing page for very large offline jobs.

Caching pays off when many requests share the same long prefix: a system prompt, tool definitions, a codebase or a reference document. The first request writes that prefix to the cache at the cache-write price. Later requests within the cache lifetime (by default a few minutes, refreshed on each hit) read it at the cache-read price, which is a small fraction of normal input.

The effect is largest for agents and coding tools, where the same context is resent on every turn, and for chatbots with long fixed instructions. It does little for one-off requests with unique content, such as the summarization job above. Output tokens are never cached, so caching cannot reduce the output side of a bill.

• Tight budget, high volume: Claude Haiku 4.5 for classification, extraction, routing and simple support replies.
• Everyday coding and most products: Sonnet (Claude Sonnet 5.5 or Claude Sonnet 4.6). It is the best value for most work.
• Hard coding and analysis: Opus (Claude Opus 5.5 and earlier Opus versions) when Sonnet's answers need too many retries.
• Highest capability: Fable (Claude Fable 5.1), the most capable and most expensive tier. Use it for the tasks that need it, not as a default.

A common pattern is to default to Sonnet and switch per task. In Claude Code you change models with /model; the same APIVAI key works for every model. If you also use GPT models, see the cheaper Claude Code and GPT API guide, or compare providers on Anthropic API alternatives.

As of 2026-10-10, most Claude models on APIVAI are 56–62% below Anthropic's list prices, depending on the model. The pricing page always shows the current figures.

Yes. Cache writes and cache reads are billed at their own prices, shown in the cache table above and on the pricing page, and each request in the dashboard shows its cache tokens.

Yes. Send a POST to /v1/messages/count_tokens with the same body you would send to /v1/messages, and the response returns the input token count.

No. You top up a prepaid balance (from $10 by card or crypto, or ¥20 via Alipay / WeChat Pay) and pay only for the tokens you use.

Generating tokens takes more compute than reading them, so every Claude model prices output at about five times its input. Keeping answers concise is one of the simplest ways to cut cost.

Create an account and get your API key →