Back to News

Anthropic launches Claude Haiku 5.5 with effort controls, 90% price cut

1045 points · 487 comments#anthropic#claude-haiku-5.5#model-release#pricing

Anthropic released Claude Haiku 5.5, its cheapest and fastest small model, priced at $0.10 per million input tokens and $0.50 per million output tokens for requests up to 100K tokens, a 90% cut from Haiku 4.5. It is the first Haiku model with adjustable reasoning effort, and benchmarks show it outperforming GPT-6 Luna on several tests, including OSWorld 2.1 and GDPval-AA v2.1.

Coverage timeline

  1. Hacker Newssfkgtbor
  2. Techmeme

    Anthropic : Anthropic launches Claude Haiku 5.5, the first Haiku model with effort controls, for high-volume, cost-sensitive tasks like summaries and classification — Introducing Claude Haiku 5.5: the cheapest, fastest, and most capable small model we've ever released. Claude Haiku 5.5 is designed for high-volume, cost-sensitive tasks.

  3. Techmeme

    Frederic Lardinois / The New Stack : Anthropic prices Haiku 5.5 at $0.10/1M input and $0.50/1M output tokens for requests up to 100K tokens, and $0.50 and $2.50 above, vs. $1 and $5 for Haiku 4.5 — Anthropic on Wednesday launched Claude Haiku 5.5, the first new version of its smallest and most affordable model in nearly a year.

  4. Simon Willison

    As previously promised , here's Anthropic's new fast, low cost model: Introducing Claude Haiku 5.5 . The previous Haiku, 4.5, was very much showing its age. It came out almost a year ago , and was priced at $1/million input and $5/million output - relatively expensive even back then, and a full 10x the price of OpenAI's GPT-6 Luna , released last month. The new Haiku exactly matches the price of GPT-6 Luna - $0.10/$0.50 - up to 100,000 tokens. Beyond 100,000 tokens the price increases 5x to $0.50/$2.50. Luna itself has a price increase at 272,000 tokens but only to $0.20/$0.75. Haiku 5.5 also uses a new, less generous tokenizer. My Claude Token Counter tool shows that the same long prompt uses around 1.25x as many tokens with Haiku 5.5 compared to Haiku 4.5, so there's a hidden price increase there. If your workloads fit in 100,000 tokens, Haiku is the same price as Luna and reports higher benchmark scores. Above 100,000 tokens, Luna looks like a much better deal. The most recent relea

  5. 量子位量子位

    Claude新模型发布!跑分暴击GPT-6 Luna,价格比梁文谷还便宜,OpenAI只能送重置卡挽尊 梦晨 2026-10-08 09:04:00 来源: 量子位 小模型新守门员 梦晨 发自 凹非寺 量子位 | 公众号 QbitAI Claude Haiku 5.5发布,小模型大逆袭! 单论跑分,直接超越两大前“斩杀线”DeepSeek V4.1Flash和GLM-5.3-Flash,成为新的小模型守门员。 一看官方跑分,好嘛,这就是冲着GPT-6 Luna来的,逐项赢过Luna。 由于鹈鹕骑自行车测试已经基本饱和了,最新干到战场的是奔跑斑马测试。 Haiku 5.5甚至价格也完全对标GPT-6 Luna。 上代Haiku 4.5还是刚好一年前出的,价格是5.5的10倍。 不过Haiku5.5的价格还有条件,提示词在10万token以内的请求(占Haiku 4.5请求量的90%),输入输出价格直接砍90%;超过10万token的部分,只砍50%。 这样一来,除了缓存命中的输入这一项,Haiku 5.5的价格也比DeepSeek-V4.1 Flash便宜了。 至此Opus 5.5管复杂推理、Sonnet 5.5管通用执行、Haiku 5.5管跑量和速度,群贤毕至,只差Fable了。 只能说隔壁OpenAI要是不努把力,Claude也没必要把Fable 5.5这管大牙膏挤出来。 仿佛回到了英特尔和AMD比拼CPU的日子。 换了tokenizer,耗费token更多 Haiku 5.5是第一个支持effort调节的Haiku模型,继承了系列从low到max的五档配置。 上一代Haiku 4.5在计算机操作和编码方面基本交白卷,这一代直接进化。 看OSWorld 2.1(测agent操作真实电脑完成多步任务):Low档42.0%准确率、单次成本$0.07;Max档72.4%、$0.61。Haiku 4.5只有Max档15.7%、$1.45, 也就是说,5.5的Low档比4.5的Max档准,还便宜一半。 GDPval-AA v2.1(测44个职业的真实专业工作)也是类似曲线:Low档Elo 1125、$0.01;Max档1620、$0.87。4.5的Max档只有735、$0.24。 不过Haiku 5.5换了tokenizer(和Sonnet 5.5、Opus 5.5同款)

  6. 机器之心机器之心

    编辑|勺嘴鹬 Anthropic 争夺 Agent 成本。 同志们,Anthropic 又来卷价格了。这次轮到 Claude 家族里最便宜的 Haiku。 10月7日,Anthropic 正式发布 Claude Haiku 5.5 ,号称自家迄今速度最快、能力最强、价格最低的小模型。 最夸张的是价格, 每百万输入Token 最低只要 0.1美元,相比上代直接降价90%。 性能也没落下。按照 Anthropic 公布的测试结果,Haiku 5.5 在计算机操作、知识工作、Agent 编程等多个基准上超过 OpenAI 的 GPT-6 Luna,部分成绩甚至拉开了不小的差距。 而且,Haiku 5.5 还引入了可调节的推理强度,支持 100万Token 上下文窗口和最高 12.8万Token 输出。 至此,Anthropic 在 短短 15 天里完成了 Claude 5.5 全家桶的更新 :9 月 22 日 Opus 5.5 登场,9 月 28 日 Sonnet 5.5 发布,如今 Haiku 5.5 也正式加入。 这次,Anthropic 把重点放在了那些 调用量巨大、对价格极其敏感的 AI 任务 上。 多项成绩超过 GPT-6 Luna,计算机操作成功率 72.4% 按照 Anthropic 公布的 Benchmark 成绩,Haiku 5.5 相比上一代实现了大幅提升,在部分测试中还取得了比 GPT-6 Luna 更好的成绩。 其中最显眼的成绩来自 OSWorld。这是一项衡量 AI Agent 操作真实计算机能力的测试,要求模型完成跨应用、多步骤的操作任务。 在 OSWorld 2.1 离线任务子集上,Haiku 5.5 取得 72.4% 的成功率,上一代 Haiku 4.5 只有 15.7%,GPT-6 Luna 则为 48.9%。 知识工作方面,Haiku 5.5 在 GDPval-AA v2.1 中取得 1620 分,超过 GPT-6 Luna的1437分。Agent 编程测试 Terminal-Bench 4.0 中,两者成绩分别为 39.2% 和 16.4%。 不过,这些数据还有一个重要细节。 Haiku 5.5 的最高 Benchmark 成绩,往往需要付出更多推理计算。 Haiku 5.5 首次在 Haiku 系列中引入可调节推理强度,开发者可以通

Related stories

  • GPT-6.2 Astra Gray-Tests for Pro Users; Claude Mythos 5.5 Leak Surfaces

    On November 2, 2026, OpenAI began gray-scale testing GPT-6.2 Astra for some Pro subscribers, with reports of 20% or even 100% of requests routed to the new model. Meanwhile, Anthropic's Claude Mythos 5.5 briefly appeared in frontend code before being removed, and a suspected leaked weight repository on HuggingFace attracted over 5,000 downloads.

    1 reports · 1 sources
  • Karpathy builds HackerTube, a YouTube clone from Hacker News video recommendations

    Andrej Karpathy announced he used Claude Opus 5.5 to build HackerTube, a YouTube clone that surfaces videos recommended in Hacker News comments, as a reaction to YouTube's algorithm. The tool processed Hacker News comments, filtered them with Sonnet, and compiled 6,454 videos recommended from 2007-2026 across 150 topics.

    1 reports · 1 sources
  • Leaked Azure entries hint at Anthropic's Claude Fable 5.5 and next-gen Opus

    Screenshots from Azure's model catalog and user tests suggest Anthropic is preparing two new Claude models: Fable 5.5 and a next-generation Opus, possibly launching as soon as next week. The Fable 5.5 entry briefly appeared in Azure's catalog, and benchmark tests show a model codenamed Obsidian 3 scoring slightly above the current Opus 5.5. The leaks come amid Anthropic's upcoming IPO.

    1 reports · 1 sources
  • Artificial Analysis launches Intelligence Index v5 with private coding and science benchmarks; Claude Opus 5.5 leads at 58

    Artificial Analysis released the Intelligence Index v5, adding two new benchmarks and three upgrades, with 60% of the index now based on private test sets or held-out answers. Claude Opus 5.5 (max) leads the composite measure with a score of 58, reflecting model capabilities across real-world tasks.

    2 reports · 1 sources

News updates every hour. Subscribe in your RSS reader to get new stories automatically.

Subscribe