Claude Haiku 5.5 Is LIVE (RIP OpenAI)

Summary
Chase from the YouTube channel Chase AI reviews Anthropic’s launch announcement and documentation page for Claude Haiku 5.5, released on October 7, 2026. He breaks down the model’s new pricing structure, effort-level scaling, safety guardrails, and benchmark comparisons against Haiku 4.5, GPT-6 Luna, and Sonnet 5.5.

What is shown

• [00:04] The official Anthropic announcement page titled “Claude Haiku 5.5” (dated October 7, 2026).
• [00:11] The introductory overview outlining Claude Haiku 5.5’s positioning as Anthropic’s fastest and cheapest small model for high-volume, agentic tasks.
• [00:43] Documentation showing a price cut for Claude Sonnet 5.5 cache reads (halved to make agentic workflows ~20% cheaper).
• [01:06] Benchmark comparison table showing GDPval-AA v2.1, Computer use, Humanity’s Last Exam, Terminal-Bench 4.0, FrontierCode 1.1, and Visual reasoning across Haiku 5.5, Haiku 4.5, GPT-6 Luna, and Sonnet 5.5.
• [02:07] Effort-level performance vs. cost graphs across OSWorld 2.1 (offline subset), GDPval-AA v2.1, Humanity’s Last Exam, and Terminal-Bench 4.0, showcasing scores across Low, Medium, High, and Max effort settings.
• [04:02] Full API pricing comparison table detailing prompt tiers (≤100k vs. >100k tokens), cache reads/writes, and input/output tokens across Haiku 5.5, Haiku 4.5, and Sonnet 5.5.
• [04:46] Safety and availability sections detailing alignment, cybersecurity safeguards, biology restrictions, and platform rollout.

Claims & numbers

• Pricing & Efficiency:
  • Claude Haiku 5.5 runs on average ~75% cheaper than Haiku 4.5 (the presenter quotes the post).
  • Claude Sonnet 5.5 cache read prices are halved, running ~20% cheaper on most agentic tasks.
  • Haiku 5.5 API pricing per million tokens (prompts ≤100k / >100k tokens): Cache reads $0.01 / $0.05 (vs. $0.10 on Haiku 4.5; $0.15 on Sonnet 5.5); Cache writes $0.125 / $0.625 (vs. $1.25 on Haiku 4.5; $2.50 on Sonnet 5.5); Input tokens $0.10 / $0.50 (vs. $1.00 on Haiku 4.5; $2.00 on Sonnet 5.5); Output tokens $0.50 / $2.50 (vs. $5.00 on Haiku 4.5; $10.00 on Sonnet 5.5).
• Benchmark Scores:
  • Knowledge work (GDPval-AA v2.1): Haiku 5.5 scores 1620 vs. 735 (Haiku 4.5), 1437 (GPT-6 Luna), and 1836 (Sonnet 5.5).
  • Computer use: Haiku 5.5 scores 72.4% vs. 15.7% (Haiku 4.5), 48.9% (GPT-6 Luna), and 83.9% (Sonnet 5.5).
  • Multidisciplinary reasoning (Humanity’s Last Exam): Haiku 5.5 scores 45.9% vs. 10.2% (Haiku 4.5) and 56.9% (Sonnet 5.5).
  • Agentic coding (Terminal-Bench 4.0): Haiku 5.5 scores 39.2% vs. 0.0% (Haiku 4.5), 16.4% (GPT-6 Luna), and 70.6% (Sonnet 5.5).
  • FrontierCode 1.1: Haiku 5.5 scores 46.4% vs. 42.4% (GPT-6 Luna) and 64.9% (Sonnet 5.5).
• Effort Levels & Cost Tradeoffs:
  • Haiku 5.5 is the first Haiku model with an adjustable effort setting (Low to Max).
  • On OSWorld 2.1, Haiku 5.5 at Max effort scores 72.4% at $0.61/attempt (matching Sonnet 5.5 on High at $1.38/attempt for under half the cost); Haiku 5.5 at Medium effort scores 55.6% at $0.13, outperforming GPT-6 Luna Max (48.9% at $0.20).
  • On Terminal-Bench 4.0, Haiku 5.5 at Max scores 39.2% at $2.64, comparable to Sonnet 5.5 on High (43.0% at $1.97).
• Safety:
  • Haiku 5.5’s cybersecurity safeguards are more restrictive than Haiku 4.5’s, but less restrictive than Sonnet 5.5’s or Opus 5.5’s.
  • Biology safeguards remain identical to Sonnet 5, Sonnet 5.5, and Opus 5.

Notable quotes

• [00:15] “Haiku 5.5 is on average 75% cheaper than Haiku 4.5.”
• [02:32] “That’s essentially the same score as Sonnet 5.5 on High, but significantly cheaper, about half the cost.”
• [05:17] “Haiku 5.5 is available everywhere right now, and I think this is a huge update, especially if you’re someone who is usage or API cost conscious.”

Assessment
This is a third-party commentary and documentation walkthrough rather than a hands-on product demonstration. The presenter displays and reads directly from Anthropic’s published launch materials, graphs, and pricing tables, noting that live real-world testing will be covered in a subsequent video.

Described by gemini-3.8-flash on 2026-10-08 from the video’s audio and frames.