Anthropic vs OpenAI

At a glance

1. Two sheets, one identical sticker

The flagship rows of both price sheets match to the cent, and so do the cache reads.

Anthropic's Opus 5.5 sits one rung up at $4 and $20, with cache reads at $0.20. OpenAI's GPT-6 Astra sits at $10 and $50. On the flagship row there is nothing to choose, which is the first sign that the model price is not the decision.

2. The tokenizer changes the invoice before your prompt does

Anthropic's pricing page carries a note most price sheets would leave out: Claude 4.7 and later use a newer tokenizer that produces approximately 30% more tokens for the same text.

So the sticker is $2 per million tokens, and the bill is $2 per million tokens of Claude's counting. For text that costs $2.00 per million source tokens on a standard tokenizer, the same text costs $2.00 × 1.3 = $2.60.

That is a 30% difference on the largest line of a RAG or agent bill, and it does not appear anywhere in a side-by-side table that compares per-token prices. The note also says the exact increase depends on the content, so treat 1.3 as a working multiplier, not a constant.

3. Caching is a break-even test, not a discount

Both vendors sell the same shape: pay more to write a prefix, pay less to read it. The multipliers are where they part.

A cached prefix costs 1.25 + 0.1n to serve n requests, against n uncached. The break-even is n = 1.39, so the second request already wins, and ten requests cost 2.25 instead of 10, a 77.5% cut on the input line. The long windows are worse deals for short runs: 2x + 0.1n beats n only from the third request, and OpenAI's 2.5x write needs three reads before it pays.

4. Batch is the one line where both sheets agree

Both vendors halve input and output for asynchronous work. Anthropic publishes the batch rows separately; OpenAI prints a Batch table and a separate Flex table at the same numbers.

Anthropic's page says its cache multipliers stack with the batch discount. That combination is the cheapest published way to process a large, stable corpus: a cached prefix at 0.1x inside a batch, on a job that can wait. The trap is that a nightly single-run job has nothing to read from, and a cold cache write just adds 25% to a batch you could have submitted uncached.

5. Long prompts are where the sheets stop being comparable

Both vendors reprice the prompt once it gets long, and they do it in different shapes.

OpenAI's long-context row doubles input and multiplies output by 1.5. Anthropic's Haiku sheet lists two rows, one for prompts up to 100,000 tokens and one for prompts over it, at five times the rate, and it does not print a marginal split.

A 200,000-token prompt therefore bills at 0.2 × $0.50 = $0.10 on the over-100k row against $0.02 if the same prompt fit under the line. That cliff, not the per-token sticker, is what a document pipeline hits when someone pastes a whole contract set.

6. Three add-ons that quietly change the unit

Fast mode on both sides doubles the sticker: Opus 5.5 goes to $8 input and $40 output, GPT-6.1 Sol to $4 and $20. Anthropic notes that fast mode is not available with the Batch API, so the two savings cannot be stacked.

Data residency is a multiplier rather than a tier. Sending Claude 4.6 or later to US-only inference applies 1.1x to every token category, cache writes and reads included.

The third is the cache window, and it is easy to lose money on. A 12-hour write costs 2.5x, so a job that runs twice a day should skip caching entirely rather than pay 2.5x for a prefix it reads once. Check the read count before enabling it: three reads is where the long window starts to earn.

Cost arithmetic at volume

These are computed from the published unit prices, not measured. Each row prints its own formula, so you can replace our assumption with your traffic and redo it.

Decision matrix

Rows are the situation you are actually in. Pick the row before you read the pricing page again.

Code and configuration

// Anthropic bills the prefix once: the breakpoint is a marker, not a hint.
import Anthropic from "@anthropic-ai/sdk";
const client = new Anthropic();

const msg = await client.messages.create({
  model: "claude-sonnet-5-5",
  max_tokens: 512,
  system: [{
    type: "text",
    text: SYSTEM_PROMPT,
    cache_control: { type: "ephemeral" }, // 1.25x write, then 0.10x reads
  }],
  messages: [{ role: "user", content: question }],
});
// Breaks when: the breakpoint sits after a timestamp or a user id. Every request
// then writes a new prefix and pays 1.25x instead of reading at 0.10x.
console.log(msg.usage);

// OpenAI caches automatically: there is no marker, only a prefix to keep stable.
import OpenAI from "openai";
const client = new OpenAI();

const res = await client.responses.create({
  model: "gpt-6.1-sol",
  input: [
    { role: "system", content: SYSTEM_PROMPT }, // byte-identical every call
    { role: "user", content: question },
  ],
});
// Breaks when: the system prompt is assembled with a date or a request id.
// A changed prefix is a cache miss billed at the full $2 per million.
console.log(res.usage?.input_tokens_details);

What you gain, what you give up

Migration path

Common questions

Sources & verification

Prices and limits were read from these pages on 2026-11-10. Vendor marketing claims are quoted as claims. Anything derived is arithmetic on the numbers below.

Keep reading