Claude Haiku 5.5 is priced like GPT-6 Luna up to 100K tokens, and five times higher beyond
Anthropic released Claude Haiku 5.5 on 7 October 2026[1,2]. The model id is claude-haiku-5-5, and it is built for high-volume, latency-sensitive tasks such as classification, extraction and routing[3].
Per million tokens, prompts of up to 100K tokens cost $0.10 input, $0.01 for cache reads, $0.125 for cache writes and $0.50 output[1]. For prompts over 100K tokens the same four rates are $0.50, $0.05, $0.625 and $2.50, five times as much each[1].
Haiku 4.5 costs $1 input and $5 output, so Haiku 5.5 is a tenth of that price up to 100K tokens and half of it beyond[1]. According to Anthropic, about 90 percent of requests to Haiku 4.5 were 100K tokens or shorter[1].
Haiku 5.5 uses a new tokenizer close to those of Sonnet 5.5 and Opus 5.5, so the same text counts as slightly more tokens than on Haiku 4.5[1,2]. With those extra tokens included, Anthropic estimates that Haiku 5.5 costs about 75 percent less to run than Haiku 4.5 on average[1].
Up to 100K tokens, all four rates are the same as OpenAI's GPT-6 Luna[1,4]. Luna's rates go up when input exceeds 272K tokens[4]. The full request is then billed at twice the input and cache rates and 1.5 times the output rate[4].
For prompts between 100K and 272K tokens, Haiku 5.5 costs five times as much as Luna on both input and output[1,4]. Past 272K, Luna rises to $0.20 input and $0.75 output, so Haiku 5.5 comes to 2.5 times Luna on input and a little over three times on output[1,4].
The context window is 1 million tokens, and one response can be up to 128K tokens[2,3]. The reliable knowledge cutoff is June 2026, and the model will not be retired before 7 October 2027[3].
The effort setting lets users choose whether to favor cost or intelligence[1]. Haiku 5.5 is the first Haiku model with this setting, and its default is medium[1,3].
On Terminal-Bench 4.0, which tests multi-step professional tasks in a command-line interface, Anthropic reports 39.2 percent for Haiku 5.5[1]. Haiku 4.5 scored 0.0 percent, GPT-6 Luna 16.4 percent and Sonnet 5.5 70.6 percent[1].
Anthropic recommends Sonnet 5.5 and Opus 5.5 for complex agentic coding, and names narrowly scoped work such as summarization and subagents as the fit for Haiku 5.5[1].