Haiku pricing: Haiku 5.5 vs 4.5 vs GPT-6 Luna

Name

Soren Atelier

Updated on 10/14/2026

Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. Above 100,000 input tokens, those rates become $0.50 and $2.50. Haiku 4.5 costs $1 input and $5 output per million tokens. These are first-party API prices in USD, checked October 14, 2026. Anthropic's pricing table (opens in a new tab) lists both Haiku 5.5 tiers.

For comparison, GPT-6 Luna has the same $0.10 / $0.50 base rates, with a higher-price tier above 272,000 input tokens. "GPT-6" is a model family, so this page names Luna explicitly. OpenAI's model page (opens in a new tab) confirms its rates and threshold.

Use the comparison chart to see how prompt length changes the ranking, or jump to cache pricing and worked costs. API pricing is usage-based. Claude app subscriptions have a separate billing system.

Haiku 5.5 API pricing at a glance

Every figure below is USD per 1 million billed tokens, at standard speed and global routing. Cache writes and reads are separate input categories.

Sources: Haiku 5.5 model specification (opens in a new tab) and Haiku 4.5 model specification (opens in a new tab). The API ID for the new model is claude-haiku-5-5.

At the same billed token counts, Haiku 5.5's input and output rates are 90% below Haiku 4.5's in the lower tier, and 50% below it in the higher tier. That is a rate comparison. It does not establish the cost of completing a task or its accuracy.

Compare the same prompt length

Select a 50k, 150k, or 300k prompt. Each selection uses that prompt length to choose the applicable rate for every model. Input, output, and cache reads have separate panels and units. Haiku 4.5 drops out at 300k because its context limit is 200k.

Here are the GPT-6 Luna rates used in the chart. OpenAI's cache-write category is listed separately from Anthropic's cache durations.

Source: OpenAI API Standard pricing (opens in a new tab), checked October 14, 2026. Batch, Flex, Fast, regional routing, and tool fees are outside this comparison.

For a 150k prompt, Haiku 5.5 uses its higher tier while Luna still uses its lower tier. Above 272k, Luna's input and cache rates double and its output rate rises by 50% for the full request. The chart compares token rates only. Different tokenizers and reasoning budgets can change the final bill.

The 100k threshold applies to the whole Haiku request

Haiku 5.5 selects its price tier from the total input of each request, including cache reads and cache writes. Once that total exceeds 100,000, the higher rates apply to the request's input categories and output. Earlier requests keep their own rates. Anthropic documents this under long-context pricing (opens in a new tab).

This creates a noticeable jump. With uncached input and 1,000 output tokens, the arithmetic is:

These are calculations from published rates, not measured API invoices. A cache hit can reduce the amount charged at the uncached-input rate, but cached tokens still count toward the threshold. For example, 90k cached tokens plus 15k new tokens makes a 105k prompt and selects the higher tier.

The 100k pricing threshold is also distinct from the context limit. Haiku 5.5 supports a 1M context window (opens in a new tab). A valid request can still cross into the higher-price tier.

Cache reads and writes have different prices

For Haiku, a 5-minute cache write costs 1.25 times the input rate. A 1-hour write costs twice that rate. A cache read costs one tenth of it. The write rate is the charge for those tokens, so do not add a second uncached-input charge for the same cached prefix. See Anthropic's prompt-caching guide (opens in a new tab).

Suppose ten requests share an 80k-token prefix. Each adds 5k uncached input tokens and produces 1k output tokens. All stay in Haiku 5.5's lower tier. If the first request writes the prefix to a 5-minute cache and the next nine get cache hits, the calculated cost is:

This example assumes actual cache hits. An expired cache or a changed prefix can require another write. In logs, count uncached input, cache creation, and cache reads separately. The prompt-caching guide (opens in a new tab) explains the usage fields.

What does a request actually cost?

For a request without caching, multiply each token count by its applicable rate:

cost in USD = input_tokens / 1,000,000 × input_rate
            + output_tokens / 1,000,000 × output_rate

For a cached Haiku request, add the cache categories at their own rates:

cost in USD = uncached_input / 1,000,000 × input_rate
            + cache_write_5m / 1,000,000 × write_5m_rate
            + cache_write_1h / 1,000,000 × write_1h_rate
            + cache_read / 1,000,000 × read_rate
            + output_tokens / 1,000,000 × output_rate

Select the Haiku 5.5 tier using the sum of all four input categories. The examples below use standard uncached requests, a fixed output count, and no paid tools.

These are equal-token scenarios, not equal-text or equal-quality benchmarks. Haiku 5.5 uses a newer tokenizer than Haiku 4.5; the migration guide (opens in a new tab) explains why the same text can produce more tokens. Use the target model's token counter and actual usage when estimating production spend. Include billed reasoning output, follow-up calls, and retries where applicable.

If you are choosing an agent for a notebook workflow, our Claude Code and Jupyter guide explains how runtime access changes the task. That decision is separate from comparing API rates.

Batch pricing is a separate comparison

Anthropic lists a 50% Batch discount on input and output. Haiku 5.5 Batch input/output prices are $0.05 / $0.25 per million tokens up to 100k, and $0.25 / $1.25 above 100k. Haiku 4.5 Batch costs $0.50 / $2.50. The 100k boundary still matters. Official Batch price table (opens in a new tab).

Compare Batch with Batch when the task can run asynchronously. A Batch quote beside a standard synchronous rate would make the cheaper service look better for the wrong reason. Check your provider's tool and routing charges separately before setting a budget.

FAQ

How much does Claude Haiku 5.5 cost?

Standard first-party API rates are $0.10 input and $0.50 output per million tokens up to 100k input tokens. Above 100k, they are $0.50 and $2.50. Cached input uses its own read or write rate.

Is Haiku 5.5 cheaper than Haiku 4.5?

At the same billed token counts, yes. Its base input/output rates are 90% lower up to 100k and 50% lower above 100k. Measure task completion and token usage before treating those percentages as workload savings.

Is Haiku 5.5 cheaper than GPT-6?

For GPT-6 Luna, the base input/output rates match. At 150k input tokens, Luna retains its lower rates while Haiku 5.5 has crossed its threshold. Other GPT-6 variants have separate prices.

Do cache hits keep Haiku 5.5 below the 100k tier?

No. Cached input counts toward total prompt length. An over-100k request uses the higher price tier even when most input comes from cache.

Related Guides

• Best LLM for coding: practical model comparisons
• Claude Code and Jupyter notebooks
• How to use DeepSeek Harness with a model API
• How to use Pi Coding Agent