Claude Haiku 4.5 pricing, $1 in and $5 out per 1M tokens

Every figure captured 6 September 2026 from the Anthropic pricing page and cross checked the same morning. Treat it as a dated snapshot.

$1 per million input tokens and $5 per million output tokens is the Claude Haiku 4.5 API price on 6 September 2026. Cache hits cost $0.10 per million, a 5 minute cache write $1.25, a 1 hour cache write $2, and the Batch API halves the base rate to $0.50 in and $2.50 out.

The question behind most of these searches is whether Haiku is still the cheap tier now that Sonnet 5 sells for $2 and $10. It is. Haiku 4.5 costs 2x less than Sonnet 5 and 3x less than Sonnet 4.6 on both input and output, and it costs 1.25x the price of the retired Haiku 3.5.

Claude Haiku 4.5 price table

Claude Haiku 4.5 and its neighbours, USD per 1M tokens, from the Anthropic pricing page read on 6 September 2026

The cache columns follow the multipliers the same page publishes, 1.25x base input for a 5 minute write, 2x for a 1 hour write, 0.1x for a hit. The page states that caching pays off after one read for the 5 minute duration and after two reads for the 1 hour duration. The prompt caching break even calculator runs that arithmetic for any prefix size.

Batch API rates, USD per 1M tokens, same source and date

The batch cache hit column is derived, not printed. The official page says the caching multipliers stack with the Batch API discount, and the OpenRouter record for the batch variant carries a cache read price of $0.05 per million tokens, the same number.

Three worked examples

Classification, 10,000 requests

Take 10,000 requests at 400 input tokens and 20 output tokens each. That's 4,000,000 input tokens and 200,000 output tokens. Input costs 4 x $1.00, which is $4.00, and output costs 0.2 x $5.00, which is $1.00. The job costs $5.00, or $0.0005 per request. On the Batch API the same job is 4 x $0.50 plus 0.2 x $2.50, which is $2.50. The same job costs $10.00 on Sonnet 5 and $15.00 on Sonnet 4.6.

Chat product, 30 days, with and without caching

A chat product sends 2,000,000 input tokens and 500,000 output tokens per day. Without caching, input is $2.00 a day and output is $2.50 a day, so 30 days costs $135.00.

Now suppose 1,500,000 of the daily input tokens sit in a shared prefix (system prompt, tool definitions, retrieved documents) and 90 percent of requests hit the cache. Per day that's 1,350,000 hit tokens at $0.10, which is $0.14 (exactly $0.135), plus 150,000 miss tokens at $1.00, which is $0.15, plus the 500,000 never cached tokens at $1.00, which is $0.50, plus output at $2.50.

The daily total is $3.29 (exactly $3.285) and 30 days costs $98.55, a saving of $36.45 or 27 percent. That figure bills the misses at base input. If every miss is instead re-written as a 5 minute cache write at $1.25, add $0.04 a day, $1.12 over 30 days, for a total of $99.68. A single 5 minute write of the whole 1,500,000 token prefix would cost $1.88. The same uncached workload costs $270.00 on Sonnet 5 and $405.00 on Sonnet 4.6.

The same 30 days on the Batch API

Move the chat workload to batch rates. Without caching, a day is 2 x $0.50 plus 0.5 x $2.50, which is $2.25, so 30 days costs $67.50. With the same 90 percent hit pattern and the batch cache hit at $0.05, a day is $0.25 plus $0.075 plus $0.0675 plus $1.25, which is $1.64 (exactly $1.6425), and 30 days costs $49.28. That's 63.5 percent below the $135.00 standard uncached figure. Batch only fits work that can wait for asynchronous results, so nightly summarisation rather than live chat. The batch API savings calculator applies the discount to any token mix.

What changed around Haiku 4.5

None of the three sources shows a Haiku 4.5 figure that differs from the table. What moved is the tier above it. The official page says the $2/$10 per million input/output token pricing for Claude Sonnet 5, announced at launch as introductory pricing through August 31, 2026, is now the standard price, and that the previously scheduled increase to $3/$15 per million input/output tokens on September 1, 2026 will not occur. So the gap between Haiku 4.5 and the current Sonnet is 2x, not the 3x that Sonnet 4.6 still shows.

Against its predecessor, Haiku 4.5 is 1.25x the price of Haiku 3.5 on every column, $1 against $0.80 on input and $5 against $4 on output. Haiku 3.5 is marked retired except on Bedrock and Google Cloud. The LiteLLM registry lists a deprecation date of 2026-10-15 for the claude-haiku-4-5 key. That is a registry field the official pricing page does not repeat, so treat it as unconfirmed.

Two notes on context. The official page says Claude 4.6 and later models include the full 1M token context window at standard pricing. Haiku 4.5 predates that, and both cross-check sources list it at 200,000 tokens of context with 64,000 max output tokens. The page also notes that Claude 4.7 and later models use a newer tokenizer that produces approximately 30 percent more tokens for the same text, so a per-token comparison against a later model that bills approximately 30 percent more tokens for the same text understates the gap in Haiku's favour. Our Claude API pricing page covers the whole line, including Opus.

Related pages