Claude Haiku 5.5 pricing

Haiku 5.5 rates on each side of 100K

These are Claude API list prices in USD per million tokens, verified October 8, 2026. The official model overview lists claude-haiku-5-5 as Active, released October 7, with a 1M-token context window and up to 128K standard output tokens.

Exactly 100,000 prompt tokens uses the lower band. At 100,001, the higher rates apply to the whole request, including output; this is not a surcharge on just the extra token. The 1M context window is a capacity limit, not a promise of the lower rate throughout that window. See Anthropic's pricing table.

Count the full prompt before choosing the band. With caching, add input_tokens, cache_read_input_tokens and cache_creation_input_tokens. A cache hit changes its billing category but does not remove tokens from the prompt. Tool definitions and conversation history also consume input tokens.

Boundary examples you can reproduce

Each row is one request with uncached input and 1,000 billable output tokens, using global routing. Output length stays fixed so you can see the effect of crossing the prompt boundary.

At the boundary, Standard cost is (100,000 × $0.10 + 1,000 × $0.50) ÷ 1,000,000 = $0.0105. The next input token makes it (100,001 × $0.50 + 1,000 × $2.50) ÷ 1,000,000 = $0.0525005.

For volume planning, 100,000 separate short requests, each using 3,000 uncached input and 500 output tokens, cost $55. The band is selected per request, so 300 million aggregate monthly input tokens do not trigger long-context rates by themselves.

Cache reads, writes and Batch are different charges

Bill each input token once at its category's rate. For an eligible short prompt, writing 10,000 tokens to a 5-minute cache and generating 2,000 output tokens costs $0.00225. A later hit reading those 10,000 tokens with the same output costs $0.0011, assuming the cache remains valid. The write is a paid setup cost.

For a longer request with 120,000 cache-read tokens, 1,000 fresh input tokens and 2,000 output tokens, the total prompt is 121,000. Its cost is (120,000 × $0.05 + 1,000 × $0.50 + 2,000 × $2.50) ÷ 1,000,000 = $0.0115. Counting only fresh input would select the wrong band.

The Batch API halves token rates for asynchronous work, and the documented cache multipliers stack with that discount. For example, the short 5-minute write above costs $0.001125 in Batch. Batch cache hits depend on whether matching requests reuse the cache while it remains valid; budget for writes and misses too. Check the prompt-caching guide for eligibility and lifetime rules.

The calculators on this site apply the prompt band to uncached input, cache reads and output. They exclude cache-write fees, Batch discounts, paid server tools, taxes and residency premiums. A cache-hit estimate assumes the cache is already populated. US-only inference adds a documented 10% token-price premium; partner cloud invoices may use their own rates.

The new tokenizer can change your budget

Anthropic's Haiku 5.5 changes guide says the same text produces approximately 30% more tokens than on Haiku 4.5. The increase depends on the content. Recount prompts with the target model rather than carrying over token counts from an older tokenizer.

As an illustration, a prompt counted as 80,000 tokens on Haiku 4.5 would become 104,000 if it increased by exactly 30%. That would cross Haiku 5.5's 100K boundary. This is arithmetic, not a measured token count or a forecast for every prompt. Leave room for history and tools when setting a prompt budget.

Adaptive thinking is enabled by default. Include billable thinking output when estimating cost, even when your application displays only the final answer. The migration guide explains effort settings and request changes, including replacing manual thinking budgets and omitting non-default sampling parameters.

API cost and subscription allowances

Claude subscriptions and Claude API usage have separate billing. Anthropic's subscription help page says Max and Team plans include monthly API credits, but you still need Console access and the API's billing rules still apply. A subscription allowance cannot be converted into guaranteed API calls by dividing its monthly fee by the table above.

A community comparison of a voxel-building task reports API-equivalent costs and says the models ran through subscriptions. Those figures are estimates at API list prices, not actual API invoices. Prompt growth and repeated calls belong in a cost estimate. One task does not establish a universal model ranking.

Compare the full cost of accepted results: every call, retry, cache write and billable thinking token. Keep the task and acceptance criteria fixed, then record completion quality and latency alongside spend. Cheaper tokens can still produce a more expensive task if it takes more turns.

Estimate, compare and check access

For Google’s alternative, compare Haiku 5.5 vs Gemini 3.8 Flash, including promotional prices, cache storage and the January 2027 rate change.

Choosing between small and flagship models? Read Haiku 5.5 vs Opus 5.5: costs and task selection for a two-model budget comparison and escalation examples.

Open the 100K example in the cost calculator · Compare the model catalog · Open Claude Console

Claude Haiku 5.5 pricing FAQ

The task-cost question comes from a community comparison; the subscription question is answered by Anthropic's help page. The boundary check follows the official price table.

Compare Claude Haiku 5.5 with other models