Anthropic released Claude Haiku 5.5, which it describes as its lightest model, on October 7 local time. To assess the company’s claim that it suits work prioritizing cost and speed, ITmedia compared Haiku 5.5 with Claude Sonnet 5.5 and GPT-6 Luna on three Japanese business-writing tasks. The test reflects conditions as of October 9; results may change with future updates.

The tasks were writing an apology email for a delayed delivery, turning rough notes into meeting minutes, and summarizing, organizing key points from, and replying to an English email. The models used only information in the prompts, without tools or web searches. ITmedia set each model’s reasoning level to “medium” and “max” and generated each response three times, producing 54 answers. Claude Fable 5.1 from Anthropic and GPT-6 Astra from OpenAI scored the answers out of 10 without seeing the model names or settings; the average score was the evaluation measure. Scores of 9–10 meant an answer could be used as is, 7–8 indicated limited edits, and 5–6 indicated substantial edits. Time was measured from CLI launch to completion. Costs were calculated by multiplying actual token use by API prices and converting at ¥158 per dollar.

For the apology email, medium scored 8.8 for GPT-6 Luna, 8.3 for Sonnet 5.5, and 8.0 for Haiku 5.5. Haiku met the requested character count in all three attempts. Per response, medium cost ¥0.13 for Haiku, ¥1.42 for Sonnet, and ¥0.09 for Luna. Haiku’s max score fell to 7.8 and took 123.8 seconds and ¥2.05—about 13 times as long and 15 times as costly as medium. Sonnet’s max scored 9.5 but took 271.3 seconds and ¥61.70 per response. Haiku’s medium answer added a reply deadline the prompt did not request; its max answer listed an October 31 delivery described as “expected” as though it were confirmed. Astra rated the latter a significant error and gave it 6 points.

For meeting minutes, Haiku medium scored 7.2, tying Sonnet. Haiku took 11.6 seconds and cost ¥0.18, compared with Sonnet’s 16.4 seconds and ¥3.22; Haiku cost about one-eighteenth as much. Luna misidentified the person responsible in medium and scored 5.50, rising to 7.3 in max. Haiku max also scored 7.3, but took 104.2 seconds and cost ¥1.90—about five times Luna max’s time and eight times its cost.

On the English email task, Haiku medium scored 6.3 and confused the parties’ commercial positions in two of three attempts. Max scored 8.2 and made no such error in any of the three attempts, but took 254.1 seconds and cost ¥4.68 per response. Sonnet medium scored 9.0 in 25.3 seconds at ¥4.69. Luna included information absent from the source email in all three medium attempts and scored in the 6-point range in both settings. It was fastest and cheapest in both: medium took 7.9 seconds and cost ¥0.11; max took 12.7 seconds and cost ¥0.18.

Across the three tasks, Haiku max averaged 160.7 seconds and ¥2.88, respectively 12.6 times the generation time and 15.4 times the cost of medium. Sonnet medium averaged 16.6 seconds, ¥3.11, and 8.2 points; ITmedia said it was faster and scored higher than Haiku max. Haiku medium cost about one-seventeenth as much as Sonnet medium and scored nearly as well on routine documents. The test was limited to three tasks, with three runs per task.

In Artificial Analysis’s third-party Intelligence Index, Haiku 5.5 max scored 43 points, up 26 points from the previous-generation Claude Haiku 4.5 and above GPT-6 Luna’s 38. According to Artificial Analysis, Haiku max generated about 162,000 tokens per task, roughly three times Luna max’s approximately 50,000. For prompts of up to 100,000 tokens, Anthropic’s API prices are $0.10 per million input tokens and $0.50 per million output tokens. Those rates are one-tenth of Haiku 4.5’s and the same as Luna’s, and one-twentieth of Sonnet 5.5’s $2 input and $10 output prices. Above 100,000 tokens, the rates are $0.50 for input and $2.50 for output. Anthropic positions Haiku 5.5 for narrow tasks such as summarization, context compression, and subagents; it says Sonnet 5.5 or Claude Opus 5.5 remains better suited to complex agentic coding.

Anthropic released Claude Haiku 5.5, which it describes as its lightest model, on October 7 local time. To assess the company’s claim that it suits work prioritizing cost and speed, ITmedia compared Haiku 5.5 with Claude Sonnet 5.5 and GPT-6 Luna on three Japanese business-writing tasks. The test reflects conditions as of October 9; results may change with future updates.

The tasks were writing an apology email for a delayed delivery, turning rough notes into meeting minutes, and summarizing, organizing key points from, and replying to an English email. The models used only information in the prompts, without tools or web searches. ITmedia set each model’s reasoning level to “medium” and “max” and generated each response three times, producing 54 answers. Claude Fable 5.1 from Anthropic and GPT-6 Astra from OpenAI scored the answers out of 10 without seeing the model names or settings; the average score was the evaluation measure. Scores of 9–10 meant an answer could be used as is, 7–8 indicated limited edits, and 5–6 indicated substantial edits. Time was measured from CLI launch to completion. Costs were calculated by multiplying actual token use by API prices and converting at ¥158 per dollar.

For the apology email, medium scored 8.8 for GPT-6 Luna, 8.3 for Sonnet 5.5, and 8.0 for Haiku 5.5. Haiku met the requested character count in all three attempts. Per response, medium cost ¥0.13 for Haiku, ¥1.42 for Sonnet, and ¥0.09 for Luna. Haiku’s max score fell to 7.8 and took 123.8 seconds and ¥2.05—about 13 times as long and 15 times as costly as medium. Sonnet’s max scored 9.5 but took 271.3 seconds and ¥61.70 per response. Haiku’s medium answer added a reply deadline the prompt did not request; its max answer listed an October 31 delivery described as “expected” as though it were confirmed. Astra rated the latter a significant error and gave it 6 points.

For meeting minutes, Haiku medium scored 7.2, tying Sonnet. Haiku took 11.6 seconds and cost ¥0.18, compared with Sonnet’s 16.4 seconds and ¥3.22; Haiku cost about one-eighteenth as much. Luna misidentified the person responsible in medium and scored 5.50, rising to 7.3 in max. Haiku max also scored 7.3, but took 104.2 seconds and cost ¥1.90—about five times Luna max’s time and eight times its cost.

On the English email task, Haiku medium scored 6.3 and confused the parties’ commercial positions in two of three attempts. Max scored 8.2 and made no such error in any of the three attempts, but took 254.1 seconds and cost ¥4.68 per response. Sonnet medium scored 9.0 in 25.3 seconds at ¥4.69. Luna included information absent from the source email in all three medium attempts and scored in the 6-point range in both settings. It was fastest and cheapest in both: medium took 7.9 seconds and cost ¥0.11; max took 12.7 seconds and cost ¥0.18.

Across the three tasks, Haiku max averaged 160.7 seconds and ¥2.88, respectively 12.6 times the generation time and 15.4 times the cost of medium. Sonnet medium averaged 16.6 seconds, ¥3.11, and 8.2 points; ITmedia said it was faster and scored higher than Haiku max. Haiku medium cost about one-seventeenth as much as Sonnet medium and scored nearly as well on routine documents. The test was limited to three tasks, with three runs per task.

In Artificial Analysis’s third-party Intelligence Index, Haiku 5.5 max scored 43 points, up 26 points from the previous-generation Claude Haiku 4.5 and above GPT-6 Luna’s 38. According to Artificial Analysis, Haiku max generated about 162,000 tokens per task, roughly three times Luna max’s approximately 50,000. For prompts of up to 100,000 tokens, Anthropic’s API prices are $0.10 per million input tokens and $0.50 per million output tokens. Those rates are one-tenth of Haiku 4.5’s and the same as Luna’s, and one-twentieth of Sonnet 5.5’s $2 input and $10 output prices. Above 100,000 tokens, the rates are $0.50 for input and $2.50 for output. Anthropic positions Haiku 5.5 for narrow tasks such as summarization, context compression, and subagents; it says Sonnet 5.5 or Claude Opus 5.5 remains better suited to complex agentic coding.