Claude Haiku 5.5 Understanding

Claude Haiku 5.5 (claude-haiku-5-5, released 7 Oct 2026) is Anthropic's fastest and cheapest tier, built for high-volume, latency-sensitive work such as classification, extraction and routing. OpenTryOn exposes it through the understand service with ClaudeUnderstandAdapter (first-party Messages API, anthropic SDK).

Good fits in fashion pipelines: tagging garment attributes, reading care labels and size tags, writing alt-text and product copy, and cheap QC passes over generated try-on images.

Property	Value
Model id	`claude-haiku-5-5`
Input / output	Text and images in, text out (no video)
Context / max output	1M tokens / 128K tokens
Price	From $0.10 / $0.50 per MTok (input / output), 50% off on Batch
Thinking	Adaptive, on by default; `effort` low / medium (default) / high / xhigh / max

Authentication

export ANTHROPIC_API_KEY="sk-ant-..."

This is the same key the Studio chat planner already uses when OPENTRYON_AGENT_LLM_PROVIDER=anthropic.

CLI

opentryon understand --model claude-haiku-5-5 \
  --image garment.jpg --prompt "List fabric, colour, fit and neckline as JSON."

# Cheapest/fastest: low effort, thinking off (thinking can only be disabled at high effort or below)
opentryon understand --model claude-haiku-5-5 --image label.jpg \
  --prompt "Transcribe the care label." --effort low --no-thinking

# Text-only also works
opentryon understand --model claude-haiku-5-5 --prompt "Write a 20-word product blurb for a linen shirt."

Flag	Notes
`--image`, `-i`	Optional image (path or URL; URLs are passed through, files are base64-encoded)
`--prompt`, `-p`	Question or instruction
`--system`	Optional system prompt
`--effort`	`low`, `medium`, `high`, `xhigh`, `max` (API default `medium`)
`--no-thinking`	Sends `thinking: {"type": "disabled"}`; rejected at `xhigh` / `max`
`--max-tokens`	Output cap, including thinking (default 4096)

temperature, top_p and top_k are deliberately not exposed: Haiku 5.5 returns HTTP 400 for any non-default value.

Python

from tryon.api.claude import ClaudeUnderstandAdapter

adapter = ClaudeUnderstandAdapter()  # reads ANTHROPIC_API_KEY
result = adapter.understand(
    image="garment.jpg",
    prompt="Describe the outfit in one sentence.",
    effort="low",
    thinking=False,
)
print(result["text"], result["usage"])

understand_image(image, ...) accepts a list of images for multi-image comparison.

Notes

MCP tool: understand_claude_haiku_5_5 (generated from the registry).
Planner: claude haiku 5.5 / haiku 5.5 / claude-haiku-5-5 pin this model. The understand default stays kimi-k2.6.
Haiku 5.5 uses the newer tokenizer (about 30% more tokens than Haiku 4.5 for the same text) — budget --max-tokens accordingly.
Cloud-only: Anthropic publishes no open weights.

Authentication​

CLI​

Python​

Notes​

Authentication

CLI

Python

Notes