Claude Haiku 5.5評測:砍價75%,電腦操作暴增4.6倍 | Claude Haiku 5.5 Review: 75% Cheaper, 4.6x Computer Use
By Kit 小克 | AI Tool Observer | 2026-10-16
🇹🇼 Claude Haiku 5.5評測:砍價75%,電腦操作暴增4.6倍
Claude Haiku 5.5是Anthropic最新發布的小型模型(2026/10/7),主打價格砍75%、電腦操作跑分暴增近4.6倍,小模型重度使用者值得一看。
Claude Haiku 5.5降價多少
10萬token以內的請求,輸入每百萬token0.10美元、輸出0.50美元,比Haiku 4.5便宜90%,超過10萬token費率跳5倍,仍比舊版便宜一半。Anthropic算出平均成本降了75%。
跑分暴增,但基期偏低
- OSWorld 2.1電腦操作:72.4%,前代僅15.7%。
- Humanity's Last Exam(有工具):57.4%,前代18.7%。
數字漂亮,但Haiku 4.5基期本來就低,容易衝出高倍數,實際能不能扛住agent任務,得靠真實工作流驗證。
1M上下文+自適應思考
Haiku 5.5支援100萬token上下文,輸出上限12.8萬token,是Haiku系列第一個有effort調節鈕的模型,可切低、中、高三段。已上架Claude Platform、AWS、GCP、Azure。輕量任務團隊沒理由不換,但吃重推理的工作,小模型再便宜也頂不上Sonnet。
好不好用,試了才知道。
🇺🇸 Claude Haiku 5.5 Review: 75% Cheaper, 4.6x Computer Use
Claude Haiku 5.5, released by Anthropic on October 7, 2026, cuts API pricing by up to 90% while pushing computer-use benchmarks up nearly 4.6x over its predecessor. If your stack leans on small models for classification, routing, summarization, or lightweight agent subtasks, this release is worth a closer look before you default to a bigger, pricier model.
How Much Cheaper Is Claude Haiku 5.5
For requests under 100,000 tokens, Claude Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens, a 90% cut from Haiku 4.5. Beyond 100K tokens, rates jump 5x but are still half of the previous price. Anthropic says the average cost drop across real traffic is around 75%, a move that lines up squarely against GPT-6 Luna's pricing in the ongoing small-model price war.
Big Benchmark Jumps, Low Starting Point
- OSWorld 2.1 computer-use benchmark: 72.4%, up from just 15.7% on Haiku 4.5.
- Humanity's Last Exam, with tools enabled: 57.4%, up from 18.7%.
Those numbers look dramatic, but Haiku 4.5 was genuinely weak at computer-use tasks to begin with, so a low baseline inflates the multiplier. Whether Haiku 5.5 can actually hold up through real browser-automation and form-filling agent workflows is something you'll still need to test yourself before trusting it in production.
1M Context and Adaptive Thinking
Haiku 5.5 ships with a 1 million token context window, output capped at 128K tokens, and is the first Haiku-class model with an adjustable effort setting — low, medium, or high — letting you trade off cost, latency, and reasoning depth per task. It's live now on the Claude Platform, AWS, GCP, and Azure (API ID: claude-haiku-5-5). If you're already running lightweight tasks on Haiku, there's little reason not to switch over. But for anything that genuinely needs heavy reasoning, a cheaper small model is still a small model — that's still Sonnet's job.
Only real use will tell. 好不好用,試了才知道。
Sources / 資料來源
- Anthropic: Introducing Claude Haiku 5.5
- VentureBeat: Anthropic launches Claude Haiku 5.5 with 90% API price reduction
- MarkTechPost: Anthropic Releases Claude Haiku 5.5, a Small Model With 1M Context
延伸閱讀 / Related Articles
- MAI-Base-2評測:微軟2.8兆參數模型,跑分未公開 | MAI-Base-2 Review: 2.8T Params, No Benchmarks Yet
- 流氓AI代理評測:研究員首次接觸失控代理群 | Rogue AI Agent Review: Researchers Make First Contact
- FTC調查評測:OpenAI、Anthropic遭50頁安全函查 | FTC Probe Review: OpenAI, Anthropic Face Safety Demands
AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends
留言
張貼留言