Haiku
The lightweight, fast model tier in Anthropic's Claude series, used for high-frequency, low-cost calls such as summarization, file naming and command completion.
Core Facts
Timeline (last 90 days)
Claude Code's `/model` command can switch between different models, such as the speed-focused Haiku or the more capable Opus.
In Claude Code, the Haiku model supports a context of about 256,000 tokens, while Opus and Fable reach about 1 million tokens.
Managed Agents' Agent model options include Opus, Sonnet, and Haiku, and the basic tool set typically includes bash, read, write, and grep.
The route skill evaluates tasks, assigns suitable work to Sonnet or Haiku, and then Opus reviews the results.
Claude Code's subagents are billed independently, so cheaper and faster models (such as Haiku) can be assigned to simple tasks, leaving the strongest model for main-thread reasoning.
The explorer agent built into Copilot CLI uses Haiku-class small models to explore the codebase, avoiding the use of large models for low-value tasks.
Changing the intent distribution from Zipf-skewed to uniform, the cost reduction for Sonnet and Haiku drops from 70% to 56%; if all 3000 questions are different, the reduction is only 38%.
Dynamic Workflows support configuring inference-time compute by stage, for example using Haiku in the scanning stage and Opus in the review stage.
At the time of course recording, Claude Code's default model was Opus 4.5, with Sonnet and Haiku also available as options.
In the self-built approach, the author added three things Hermes lacks: citing sources (when citing, state the original words, the decision-maker, and the date, and admit not knowing when you don't know), importing history (one command scans existing Claude Code sessions, summarizes them with Haiku and embeds them into long-term memory while retaining full transcripts), and team isolation (each memory is tagged by ownership, and queries are filtered by asker).
40 more timeline events
All Facts (20)
In Anthropic's model naming system, Opus represents the most powerful version, Sonnet is the mid-tier version, and Haiku is the lightweight version.
65%UnverifiedAnthropic's Claude model naming system borrows from musical terminology: Haiku (lightweight and fast), Sonnet (balancing performance and speed), Opus (highest performance)
90%UnverifiedAnthropic推荐使用编排器+子代理架构，用推理能力强的模型负责规划决策，用Sonnet或Haiku执行具体点击操作
90%Unverified在Anthropic的模型体系中，Opus系列定位为最高能力层级的模型，在推理深度、代码生成质量和长上下文理解方面优于Sonnet和Haiku系列
90%UnverifiedIn Anthropic's model naming system, Opus, Sonnet, and Haiku correspond to high, medium, and low tiers of capability and cost respectively.
60%UnverifiedIn Claude Code, the Haiku model supports a context of about 256,000 tokens, while Opus and Fable reach about 1 million tokens.
50%UnverifiedClaude Code's `/model` command can switch between different models, such as the speed-focused Haiku or the more capable Opus.
50%UnverifiedManaged Agents' Agent model options include Opus, Sonnet, and Haiku, and the basic tool set typically includes bash, read, write, and grep.
50%UnverifiedThe route skill evaluates tasks, assigns suitable work to Sonnet or Haiku, and then Opus reviews the results.
50%UnverifiedClaude Code's subagents are billed independently, so cheaper and faster models (such as Haiku) can be assigned to simple tasks, leaving the strongest model for main-thread reasoning.
50%UnverifiedThe explorer agent built into Copilot CLI uses Haiku-class small models to explore the codebase, avoiding the use of large models for low-value tasks.
50%UnverifiedChanging the intent distribution from Zipf-skewed to uniform, the cost reduction for Sonnet and Haiku drops from 70% to 56%; if all 3000 questions are different, the reduction is only 38%.
50%UnverifiedDynamic Workflows support configuring inference-time compute by stage, for example using Haiku in the scanning stage and Opus in the review stage.
50%UnverifiedAt the time of course recording, Claude Code's default model was Opus 4.5, with Sonnet and Haiku also available as options.
50%UnverifiedIn the self-built approach, the author added three things Hermes lacks: citing sources (when citing, state the original words, the decision-maker, and the date, and admit not knowing when you don't know), importing history (one command scans existing Claude Code sessions, summarizes them with Haiku and embeds them into long-term memory while retaining full transcripts), and team isolation (each memory is tagged by ownership, and queries are filtered by asker).
50%UnverifiedBeBox is an application running on the Haiku operating system, aiming to replicate Be Inc.'s classic hardware, the BeBox workstation, through software simulation.
50%UnverifiedAccording to discussion on Hacker News, the BeBox application is not a complete system-level emulator (such as QEMU), but more like a hardware behavior simulation layer, reproducing the BeBox's unique hardware components through Haiku's own APIs and driver framework.
50%UnverifiedThe BeBox application is released as open source.
50%UnverifiedOpus is the name of the flagship model in Anthropic's Claude series, which also includes two tiers: Sonnet (mid-weight) and Haiku (lightweight).
50%UnverifiedHaiku-level models cost about one percent of Opus per thousand tokens.
50%