Claude API Pricing (November 2026): Every Model, Prompt Caching, Batch and Real Cost Examples

If you are building anything on Claude in late 2026, the first question is always the same: what will this actually cost me? Anthropic changed a lot this autumn. Claude Opus 5.5 arrived in September at a lower price than the Opus it replaced, Sonnet 5’s introductory price became permanent, and Claude Haiku 5.5 landed at just $0.10 per million input tokens with a twist: the price jumps when your prompt is over 100,000 tokens. On top of that, cache reads on Opus 5.5 and Sonnet 5.5 now cost 5% of the input price instead of the usual 10%.

I went through Anthropic’s official pricing page line by line (checked on November 5, 2026) and turned it into one guide with every current model, every discount, the extra tool charges people miss, and worked examples so you can estimate your own bill.

Claude API Prices for Current Models

Anthropic bills the Claude API per token, with separate rates for input (what you send) and output (what Claude writes back). Prices are quoted per million tokens, written as MTok. Here is the current lineup that Anthropic recommends in its models overview, with the model ID you put in your API call.

A few things stand out. Opus 5.5 is Anthropic’s suggested starting point for most workloads, and at $4 / $20 it is 20% cheaper than Opus 5 and Opus 4.5 to 4.8, which all cost $5 / $25. Sonnet 5.5 is exactly half the price of Opus 5.5. Haiku 5.5 is the headline: for short prompts it costs one fortieth of Opus 5.5 per token, which makes it a serious option for classification, routing, extraction and high-volume chat.

There is also Claude Mythos 5.1 at the same $10 / $50 as Fable 5.1, but it is limited to organisations verified through Anthropic’s programs such as the Cyber Verification Program, so most developers will not see it in their console.

Tokenizer note: Anthropic says Claude Opus 4.7 and every newer model use a newer tokenizer that produces about 30% more tokens for the same text than Sonnet 4.6 and older. So when you compare an old Sonnet 4.6 bill with Sonnet 5.5, do not compare only the price per token. The same document will count as more tokens on the new models.

Older and Legacy Claude Model Prices

Older models stay available for a while after a new release, and plenty of production apps still pin them. These are the official prices for the ones you are most likely to still be using:

The Sonnet 5 line deserves a mention. It launched with introductory pricing of $2 / $10 through August 31, 2026, with a planned rise to $3 / $15. Anthropic cancelled that rise, so $2 / $10 is now its standard price. If you are still on Haiku 4.5 at $1 / $5, moving to Haiku 5.5 cuts your per-token price by 90% for prompts under 100K tokens.

Retired models may still run on Amazon Bedrock or Google Cloud, which set their own schedules and invoice you directly.

Prompt Caching Prices

Prompt caching is the single biggest lever on a Claude bill. If every request starts with the same long system prompt, tool list or document, you can write that prefix to a cache once and then read it back on later requests at a fraction of the normal input price. You can turn it on in two ways: add one top-level cache_control field and let the API manage breakpoints automatically (Anthropic’s recommended starting point), or place cache_control on specific content blocks yourself.

The reduced read rate is new with this generation. On Opus 5.5 and Sonnet 5.5 a cache hit costs 5% of the normal input price, so a cached 100K-token prefix on Opus 5.5 costs $0.02 per request instead of $0.40. With a 5-minute cache the write costs 1.25x once, so caching pays for itself from the very first reuse.

Caching also helps with rate limits. Anthropic’s rate limit docs say that for most models only uncached input tokens count toward your input-tokens-per-minute limit, so reads from cache let you push far more traffic through the same tier.

Caching stacks with the Batch API discount and with the data residency multiplier, so you can combine them.

Batch API: 50% Off

If you do not need an answer within seconds, the Message Batches API halves both input and output prices. You submit a file of requests, Anthropic processes them asynchronously, and you collect the results later. It is ideal for nightly summaries, bulk classification, evaluation runs, translating a catalogue or generating SEO descriptions.

One bonus that is easy to miss: on the Batches API, Opus 5.5, Sonnet 5.5, Haiku 5.5 and several earlier models can produce up to 300K output tokens per request with the matching beta header, against 128K on the normal synchronous endpoint. Fast mode cannot be combined with batch.

The Haiku 5.5 100K Price Step

Haiku 5.5 is the only current Claude model priced by prompt length. Anthropic’s rule is simple but strict:

• If the request’s prompt is 100,000 tokens or less, you pay $0.10 input and $0.50 output.
• If the prompt is over 100,000 tokens, the whole request pays $0.50 input and $2.50 output, not just the part above 100K.
• The prompt length counts all input tokens, including cache reads and cache writes.
• Each request is judged on its own, so one long request does not change the price of the next.

In practice this means you should keep Haiku 5.5 prompts under 100K wherever you can. If you run a long agent loop where the history keeps growing, trimming or summarising the history before it crosses 100K saves 80% on that request. Above the line, Haiku 5.5 still costs half of Haiku 4.5.

All other models from Claude 4.6 onwards (Opus 5.5, Sonnet 5.5, Fable 5.1 and the rest) 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 one.

Fast Mode, Web Search, Code Execution and Other Extras

Tokens are not the only line on the invoice. These are the extra charges from the official page:

Computer use and the new browser use toolsets are billed as normal tokens, but declaring them adds a fixed overhead of roughly 4,500 tokens (computer use) or 6,600 tokens (browser use) to each request, before any screenshots. Screenshots are billed as image input. If you run these agents in a long loop, prompt caching that overhead makes a big difference.

Claude is also sold through Amazon Bedrock, Google Cloud and Microsoft Foundry. Bedrock and Google Cloud (Google managed) invoice you directly at their own rates, and regional endpoints there carry a 10% premium over global endpoints.

Real Cost Examples

These are worked examples using the official rates. They are estimates from the price list, not measured bills, but they show how much the model choice and the discounts matter.

1. Customer support chatbot

30,000 conversations a month, each about 3,000 input tokens (system prompt plus history) and 400 output tokens. That is 90M input and 12M output tokens a month.

Caching the shared system prompt would lower all three further. For a simple FAQ bot, Haiku 5.5 at about $15 a month is hard to beat.

2. Document Q&A with prompt caching

An 80K-token handbook sits in the prompt of every request on Sonnet 5.5. Users ask 20,000 questions a month, each adding 500 new input tokens and getting 600 output tokens.

• Without caching: 1,610M input tokens x $2 = $3,220, plus 12M output x $10 = $120. About $3,340 a month.
• With caching: 1,600M cached tokens read at $0.10 = $160, 10M fresh input x $2 = $20, output $120, plus cache writes. If the cache is rewritten about once an hour (720 writes of 80K tokens at the 1.25x or 2x write rate), writes add roughly $144 to $230. Total about $444 to $530 a month.

That is a saving of around 85% for the same answers.

3. Bulk product descriptions with the Batch API

50,000 product descriptions on Haiku 5.5, each with 1,500 input and 300 output tokens: 75M input and 15M output.

• Standard: 75 x $0.10 + 15 x $0.50 = $15
• Batch: 75 x $0.05 + 15 x $0.25 = $7.50

4. A one-hour agent session (Anthropic’s own example)

Anthropic’s pricing page works through a one-hour Claude Managed Agents coding session on Opus 5.5 that uses 50,000 input and 15,000 output tokens: $0.20 input + $0.30 output + $0.08 runtime = $0.58. If 40,000 of those input tokens are cache reads, the total drops to $0.428.

How to Cut Your Claude API Bill

1. Route by difficulty. Send easy jobs (classification, extraction, short replies) to Haiku 5.5, everyday coding and writing to Sonnet 5.5, and keep Opus 5.5 or Fable 5.1 for the hard cases. Anthropic’s own advice is Haiku for simple tasks, Sonnet for most production workloads and Opus for the most complex reasoning.
2. Cache every repeated prefix. System prompts, tool definitions, few-shot examples and reference documents. Put the stable parts first and the changing user message last.
3. Batch anything that can wait. 50% off, and it stacks with caching.
4. Keep Haiku 5.5 prompts under 100K tokens. Summarise long histories before they cross the line.
5. Lower the effort setting. Opus 5.5 and Fable 5.1 use adaptive thinking, which you steer with the effort parameter (Opus 5.5 defaults to medium, Fable 5.1 to high). Thinking tokens are billed as output, so lower effort on simple requests saves real money.
6. Set a spend limit. In the Claude Console billing page you can set your own monthly limit below your tier’s cap so a runaway loop cannot surprise you.
7. Use global inference unless you really need US-only processing, which costs 10% more.

Claude API vs Claude Pro, Max and Ultra

The API and the Claude apps are billed separately. A Claude Pro subscription does not give you API credits. However, the higher plans now do: according to claude.com/pricing, Claude Max ($100/month) includes $100 a month in API credits and Claude Ultra ($200/month) includes $200 a month. If you are a developer who also uses the Claude app every day, Max can effectively pay for itself. I compare the plans in detail in Claude Free vs Pro vs Max vs Ultra: usage limits explained.

New API accounts get a small amount of free credit to test with. Organisations are then placed on usage tiers with monthly spend caps: $500 on Start, $1,000 on Build and $200,000 on Scale, with Custom tier above that. If you hit the cap, requests return a 429 error until the next month or until you move up a tier. I explain that error and the others in Claude API errors 429, 529 and overloaded: how to fix them.

How to Get a Claude API Key and Make Your First Call

1. Go to the Claude Console at platform.claude.com and sign up with your email or Google account.
2. Open Billing and add credit. The API is prepaid: you buy credit and usage is deducted from it. Turn on auto-reload only if you also set a monthly spend limit.
3. Open API keys, click Create key, name it after the project, and copy it once. Store it in an environment variable, never in your code or a public repo.
4. Set Limits for the workspace so a bug cannot run up a large bill overnight.

Test the key with curl (macOS/Linux; on Windows use WSL or Git Bash):

export ANTHROPIC_API_KEY="your-key-here"

curl https://api.anthropic.com/v1/messages \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{"model": "claude-haiku-5-5", "max_tokens": 200, "messages": [{"role": "user", "content": "Say hello in five words."}]}'

The same call in Python, with the official SDK (pip install anthropic). The response includes a usage block, which is exactly what you are billed for:

import anthropic

client = anthropic.Anthropic()  # reads ANTHROPIC_API_KEY

msg = client.messages.create(
    model="claude-sonnet-5-5",
    max_tokens=500,
    messages=[{"role": "user", "content": "Summarise the benefits of prompt caching in 3 bullets."}],
)
print(msg.content[0].text)
print(msg.usage)  # input_tokens, output_tokens, cache_read_input_tokens

Estimate Your Bill Before You Build

Paste your real prompt into the free token counting endpoint (client.messages.count_tokens) to see the exact input size, then plug the numbers into this small calculator. It uses the prices in the tables above.

PRICES = {  # USD per million tokens: (input, output, cache read)
    "claude-fable-5-1": (10.00, 50.00, 0.25),
    "claude-opus-5-5": (4.00, 20.00, 0.20),
    "claude-sonnet-5-5": (2.00, 10.00, 0.10),
    "claude-haiku-5-5": (0.10, 0.50, 0.01),  # prompts up to 100K tokens
}

def monthly_cost(model, requests, input_tok, output_tok, cached_tok=0, batch=False):
    inp, out, cache = PRICES[model]
    per_req = ((input_tok - cached_tok) * inp + cached_tok * cache + output_tok * out) / 1_000_000
    total = per_req * requests
    return total * 0.5 if batch else total

# 50,000 chats a month, 2,000 input tokens (1,500 cached), 300 output
print(round(monthly_cost("claude-sonnet-5-5", 50_000, 2_000, 300, cached_tok=1_500), 2))

That example comes to about $207.50 a month on Sonnet 5.5: $50 for uncached input, $7.50 for cache reads and $150 for output. Notice that output is the biggest share, so capping max_tokens and asking for short answers saves more than anything else. Cache writes cost a little extra the first time a prefix is stored, which this simple calculator leaves out.

Billing and Limit Errors You Will Hit

For the full list and retry code, see my Claude API errors guide. To compare Claude with OpenAI, Gemini, Grok and DeepSeek on the same workloads, see my AI API pricing comparison.

Frequently Asked Questions

How much does the Claude API cost per message?

It depends on length. A typical chat turn of 2,000 input and 500 output tokens costs about $0.018 on Opus 5.5, $0.009 on Sonnet 5.5 and $0.00045 on Haiku 5.5.

Is the Claude API free?

No, but new accounts receive a small amount of free credit to test with. After that you pay per token.

What is the cheapest Claude model?

Claude Haiku 5.5 at $0.10 input and $0.50 output per million tokens for prompts up to 100K tokens, or half that with the Batch API.

Do Batch and caching discounts stack?

Yes. Anthropic confirms they combine, and data residency multipliers apply on top.

Does a 1M token context cost more?

Not on Claude 4.6 and later models. The full 1M context is billed at standard per-token rates. The only exception is Haiku 5.5’s step above 100K prompt tokens.

Is Claude cheaper than GPT or Gemini?

At the flagship level, Opus 5.5 ($4 / $20) costs less than GPT-6 Astra ($10 / $50). See my AI API pricing comparison for every provider side by side, and the DeepSeek vs Claude comparison for the budget open-weights option.

How do I see how much I have spent on the Claude API?

Open the Usage and Cost pages in the Claude Console. You can break spend down by API key, model and workspace, and every API response also returns token counts in its usage field.

Is there a monthly minimum for the Claude API?

No. You only pay for tokens you use from prepaid credit. There is no subscription fee for the API itself.

Final Thoughts

Claude’s pricing in late 2026 is friendlier than it has ever been. Opus 5.5 is cheaper than the Opus models before it, Sonnet kept its $2 / $10 price, and Haiku 5.5 makes high-volume work almost free as long as you stay under 100K tokens per prompt. The real savings, though, come from how you call the API: cache what repeats, batch what can wait, and route each job to the smallest model that does it well.

Prices change, so always check Anthropic’s official page before you sign a budget. If you want to compare against the other big providers, read my OpenAI API pricing and Gemini API pricing and free tier guides. And if you would rather pay nothing per token, my best Ollama models list shows what you can run on your own GPU.

Sources

• Anthropic: Claude API pricing (official)
• Anthropic: Models overview
• Anthropic: Rate limits and spend tiers
• Claude plans and pricing