Skip to content
dalfidan.dev

← Writing

Claude API pricing and token counts: a reference (November 2026)

·6 min read·ai · llm · cost · tokens
Watercolor illustration of an open ledger book, a stack of small tokens and a brass balance scale on a wooden desk in warm light.

This page collects the Claude API prices, multipliers and token measurements behind the practical checklist. All prices are Anthropic list prices in USD per million tokens (MTok), read on 2026-11-17, unless marked as OpenAI.

Model prices

Model Input 5-min cache write 1-hour cache write Cache hit Output Batch input / output
Claude Opus 5.5 $4 $5 $8 $0.20 $20 $2 / $10
Claude Sonnet 5.5 $2 $2.50 $4 $0.10 $10 $1 / $5
Claude Haiku 5.5 (prompts up to 100k tokens) $0.10 $0.125 $0.20 $0.01 $0.50 $0.05 / $0.25
Claude Haiku 5.5 (prompts over 100k tokens) $0.50 $0.625 $1 $0.05 $2.50 $0.25 / $1.25

Output costs 5 times as much as input for each model above. For comparison, Claude Opus 4.6 lists at $5 input and $25 output.

Modifiers that change the bill:

Modifier Effect
Haiku 5.5 over 100,000 tokens The whole request pays the higher rates, cache reads included. Prompt length counts all input tokens, including cache reads and writes.
US-only inference (inference_geo: "us") 1.1x on all token categories for Claude 4.6 and later models
Fast mode on Opus 5.5 $8 input, $40 output (double the standard rate); not available with the Batch API

Prompt caching multipliers

Cache operation Multiplier Duration
5-minute cache write 1.25x base input price 5 minutes
1-hour cache write 2x base input price 1 hour
Cache read (hit) 0.1x base input price on most models Same as the preceding write
Cache read on Opus 5.5 and Sonnet 5.5 0.05x base input price Same as the preceding write

Per Anthropic, caching pays off after one cache read with the 5-minute write, and after two reads with the 1-hour write. The multipliers stack with the Batch API discount and data residency pricing.

Minimum cacheable prompt length

Minimum Models
512 tokens Opus 5.5, Sonnet 5.5, Haiku 5.5
1,024 tokens Sonnet 4.6
2,048 tokens Opus 4.7
4,096 tokens Opus 4.6, Opus 4.5, Haiku 4.5

Shorter prompts are processed without caching and no error is returned. If both cache_creation_input_tokens and cache_read_input_tokens are 0, nothing was cached.

Static tool overhead is a common caching target. Anthropic’s pricing page says the browser use toolset adds about 6,600 input tokens per request, and the computer use toolset about 4,500.

Batch API facts

Item Value
Price “50 percent of the standard API prices,” input and output
Typical completion Most batches finish within 1 hour
Expiry A batch that has not completed within 24 hours expires
Caching inside batches Discounts stack; cache hits are best effort, with typical hit rates of 30 to 98 percent
Suggested cache 1-hour cache for shared context, since batches can run longer than 5 minutes
Fast mode Not available with batch

Tokenizer measurement: Claude 4.x versus 5.5

Anthropic states that Claude 4.7 and later models use a newer tokenizer that “produces approximately 30% more tokens for the same text,” with the exact increase depending on content.

Method: on 2026-11-17, four texts were sent to Anthropic’s messages/count_tokens endpoint for six models: claude-opus-4-6, claude-sonnet-4-6, claude-haiku-4-5-20251001, claude-opus-5-5, claude-sonnet-5-5 and claude-haiku-5-5. The three 4.x models returned identical counts, and so did the three 5.5 models.

Sample text Characters Claude 4.6 / Haiku 4.5 Claude 5.5 models Increase
English prose (one paragraph, repeated 3 times) 1,752 344 503 +46%
Python code 3,432 1,458 1,603 +10%
JSON payload 3,568 1,547 1,851 +20%
Turkish prose (one paragraph, repeated 3 times) 1,629 710 851 +20%

Caveats: each count is one run and includes a small fixed message overhead. The samples are short and synthetic, so this is a spot check, not a benchmark. Only input tokens can be counted this way; output tokens are not covered.

Effect on input price, Opus 5.5 ($4) versus Opus 4.6 ($5), for the same text:

Sample Opus 5.5 cost per $5 of Opus 4.6 spend
Python code (+10%) about $4.40
JSON payload, Turkish prose (+20%) about $4.79
English prose (+46%) about $5.85

Worked example: caching a static prefix

Hypothetical workload: a 20,000-token static prefix sent with 1,000 requests to Sonnet 5.5, each within 5 minutes of the previous one.

Line Calculation Cost
Uncached 20M tokens x $2/MTok $40.00
Cache write (once) 20,000 x $2.50/MTok $0.05
Cache reads (999) 19.98M x $0.10/MTok $2.00
Cached total write + reads about $2.05

The cached prefix costs about 95 percent less. The variable part of each prompt and all output tokens are unchanged.

Worked example: output share of cost

Hypothetical Sonnet 5.5 call with 3,000 input tokens and 800 output tokens.

Part Tokens Share of tokens Cost Share of cost
Input 3,000 79% $0.006 43%
Output 800 21% $0.008 57%
Total 3,800 100% $0.014 100%

Cutting output to 400 tokens saves about 29 percent of the request. Cutting input by 1,000 tokens saves about 14 percent.

When caching saves nothing

  • Prefix used once: you pay 1.25x (or 2x) and get no read.
  • Prefix below the minimum: silently not cached (see the table above).
  • Early content changes: the cache matches an exact prefix in the order tools, system, messages. A timestamp or user ID near the top forces a new write every request.
  • Sparse traffic: the 5-minute lifetime refreshes on each hit; a longer gap means a new write.
  • Mid-stream edits: changing one level (tools, system, messages) invalidates that level and everything after it. Changing tool definitions invalidates the whole cache.

OpenAI, for comparison

OpenAI’s pricing page lists gpt-6.1-sol at $2 input and $10 output per million tokens, also a 5x ratio. On GPT-5.6 and later, OpenAI cache reads cost 0.1x uncached input (0.05x on GPT-6.1 Sol and Terra), and writes cost 1.25x for the 30-minute cache or 2.5x for the 12-hour cache. The minimum cacheable prefix is 1,024 tokens, and earlier OpenAI models have no cache-write charge. OpenAI’s Batch API also gives a 50 percent discount, with a 24-hour completion window.

Last verified: 2026-11-17 (prices and documentation), token counts measured 2026-11-17. Prices, discounts and model names change; check the linked pages before you rely on a number.

Sources