Haiku 5.5 explained in 7min..

Summary

Caleb from the channel “Caleb Writes Code” breaks down the newly released Claude Haiku 5.5 from Anthropic, evaluating its pricing structure and cost-efficiency trade-offs against competitors like OpenAI’s GPT-6 Luna. He explains why cheap, fast models are crucial for agentic workflows and demonstrates real-world computer-use and autonomous desktop tasks using his personal agent “Atticus” powered by Haiku 5.5.

What is shown

• [00:00 - 01:00] Cost-efficiency and token-efficiency graphs from Artificial Analysis comparing Claude Haiku 5.5, GPT-6 Luna, and older frontier models.
• [01:04 - 01:17] Anthropic’s announcement post demonstrating Claude 5.5 managed agents performing an egg-drop simulation test.
• [01:18 - 01:31] Demonstration of Caleb’s personal voice agent “Atticus”: Caleb asks it to analyze a research paper on his screen (H-JEPA), and Atticus automatically searches arXiv and opens related papers in browser windows in the background.
• [01:44 - 01:58] Atticus is instructed to inspect VS Code, search PyTorch documentation for scaled_dot_product_attention, and snap the documentation window to the left side of the screen.
• [02:08 - 02:37] Artificial Analysis Intelligence Index vs. Cost per Task quadrant plot comparing Haiku 5.5, GPT-6 Luna, Qwen models, and Claude 5.5 family models.
• [04:08 - 04:18] Anthropic benchmark comparison chart showing Haiku 5.5, Haiku 4.5, GPT-6 Luna, and Sonnet 5.5 scores across knowledge work, OSWorld computer use, reasoning, and coding benchmarks.
• [04:31 - 05:04] Screen recording of OSWorld 2.1 benchmark tasks running across multi-window desktop operating system environments (slide formatting, web travel booking).
• [05:05 - 05:27] Atticus executing desktop computer-use: Caleb asks it to open Microsoft Excel and toggle it from light mode to dark mode via system options UI navigation.
• [05:54 - 06:33] Haiku 5.5 attempting to browse Artificial Analysis, apply custom filters to the cost chart, and struggling before Caleb sends a follow-up refinement prompt.

Claims & numbers

• The presenter notes Anthropic has two-tiered pricing for Haiku 5.5: $0.10 per million input tokens for under 100k tokens, and $0.50 per million input tokens for context exceeding 100k tokens (a 5x increase).
• The presenter cites Artificial Analysis benchmark averages: weighted task cost of $0.21 per intelligence task for Haiku 5.5 compared to $0.07 for GPT-6 Luna.
• The presenter states that running Atticus with Haiku 5.5 or GPT-6 Luna averages roughly $0.001 per user turn.
• According to Anthropic’s benchmark table shown:
  • Computer use (OSWorld 2.1 offline subset): Claude Haiku 5.5 scored 72.4%, compared to 15.7% for Haiku 4.5, 48.9% for GPT-6 Luna, and 83.9% for Claude Sonnet 5.5.
  • Multidisciplinary reasoning (Humanity’s Last Exam with tools): Haiku 5.5 scored 57.4% vs 18.7% for Haiku 4.5 and 64.5% for Sonnet 5.5.
  • Visual reasoning (ChartQA): Haiku 5.5 scored 46.4% vs 6.4% for Haiku 4.5, 29.1% for GPT-6 Luna, and 61.6% for Sonnet 5.5.
• During the Artificial Analysis web automation test, Haiku 5.5 achieved the task in 33 seconds at 173.2 output tokens per second, though it required a prompt clarification after misidentifying chart filters.

Notable quotes

• [01:05] “According to Anthropic, Haiku 5.5 is meant for high-volume, cost-sensitive tasks.”
• [02:00] “And models like Haiku 5.5 and GPT-6 Luna are so cheap that every interaction that I have with them through Atticus averages around $0.001.”
• [04:12] “Haiku 5.5 scored 72.4% on OSWorld 2.1 computer use.”

Assessment

This is an independent creator review and hands-on testing video. The author demonstrates real, working software integrations with live voice tool-calling and desktop automation, while openly showcasing both the strengths and current limitations (such as UI filter failures and 33-second latency) of Haiku 5.5’s computer-use capabilities.

Described by gemini-3.8-flash on 2026-10-12 from the video’s audio and frames.