Haiku 5.5 ทำงานจริงได้แค่ไหน? สองโจทย์จากคลิปไทย และตัวเลขราคาที่ต้องแยกให้ออก

Anthropic unveiled Haiku 5.5 on 7th October 2026, touting it as the smallest, fastest, and best model they've ever created [1]. Pricing per token is higher than Haiku 4.5, but costs per run have dropped by an average of 75% [1]. The question remains, how well can such a small model actually perform in real-world tasks? Two examples from Thai clips and token price discrepancies have been tested [2].

The first task involves creating an online course as an admin, responding to 10 customer cases with only 8 course details [2]. The second task is to create a 5-page Thai language slide for a business starting to use AI [2]. The reported results show the first task was completed in less than a minute, while the second took around one and a half minutes [2].

The presenter confirms the model's behavior from Anthropic's documentation, though they haven't independently verified the timings and performance [2]. The first result: the model accurately answers all 10 cases in one table, separating which answers can be given and which need further action [2]. Four cases are cited in the clip [2].

Customers ask about expired coupon codes, which the model correctly identifies as invalid, but fails to provide a refund [2]. Regarding OTP codes, the model refuses to handle it [2]. The most interesting point is the second task, where the model appears unsure about trusting user-provided information [2]. The Anthropic report states Haiku 5.5 is designed for high-volume tasks requiring rapid responses [3], while the launch announcement emphasizes its ability to handle cost-intensive tasks [1].

The ad admin task matches this description, making it a good fit for the model [2]. The slide creation task, however, raises concerns as the model doesn't fully open the document as reported [2], and some pages have blank spaces that need adjustment [2]. Users should verify their own results rather than trusting model claims [2].

Specifications and scores in Anthropic's announcement should be read carefully [1]. The table comparing performance metrics shows Haiku 5.5 lags behind Sonnet 5.5 and Terminal-Bench scores [1]. Promoted use cases include customer support and browser use [1]. Token pricing differs from run pricing, with a lower per-token rate up to 100,000 tokens, but a higher rate thereafter [4].

Compared to Haiku 4.5, the per-token cost increases 90% for projects up to 100,000 tokens [4]. The reported 75% cost reduction for run-related expenses is unclear, as it refers to run costs, not token pricing [1]. The discrepancy in token counts means review of pricing needs to be recalculated [1]. Two warnings: token pricing jumps significantly after 100,000 tokens, and tokenization increases word counts by about 30% compared to Haiku 4.5 [3].

Testing suggests Haiku 5.5 is best used for tasks like high-volume customer support questions it excels at categorization and volume, but not for complex coding tasks requiring testing [1]. It's recommended to use Haiku 5.5 for preprocessing tasks before handing off to larger models [1]. I haven't personally invoked Haiku 5.5 via API or run the two clipped tasks, but based on the announcement and documentation, it appears to be a subagent designed to work alongside Opus 5.5 or Sonnet 5.5 [1].