ORBiS TREE · ENGLISH DOCUMENT
Products →Claude
A family of general-purpose AI models and conversational products developed by Anthropic.
한국어
English
This English page is a curated translation layer linked to the Korean source article. Community changes are currently made on the Korean source, where the full revision history and anonymous edit trail are preserved.
Overview
Claude is an AI assistant and model family developed by Anthropic. Its 2023 introduction described conversation, summarization, question answering, writing and coding.[1]
Training direction
Anthropic emphasizes safety and steerability in designing Claude. Its constitution explains intended behavioral principles and connects safety research with product design.[2]
Scope
The name covers evolving model generations and product experiences within the LLM ecosystem. Inspect current official materials for an individual model's performance, prices or context length rather than treat them as permanent family properties.
A model announcement and comparison conditions
On September 22, 2026, Anthropic introduced Opus 5.5 as the first Claude 5.5 model, covering coding, knowledge work, tools and safety, with detailed conditions linked through its system card.[3] This is a dated announcement, not a definition of every later offering.
The announced 40% cost reduction on typical workloads belongs to the company's tasks and settings; it differs from the same publication's input/output token-rate reduction.[3] Task success, input/output/cache usage, inference settings and latency and throughput affect comparisons. See context windows and tokenization for long-input handling.
Reading tool-use and safety materials
An agent evaluation examines actions as well as wording. The announcement discusses prompt injection, actions outside permitted scope and irreversible actions.[3] Interpret results within their tested conditions and safeguards rather than extend them to every situation.
Further reading
- Safety evaluation and red teaming explain tested behaviors and risks.
- Model cards document intended use, evaluation conditions and limitations.
- Agents explain multi-tool workflows; Anthropic covers the organization and research context.
Retain model, product and settings in comparisons
Check a specific model's properties and access path in the model overview ↗.[4] Separate the Claude family name, actual request identifier, app/API, inference settings and tools. Verify equivalent conditions before comparing app experiments with developer-interface results.
A fictional coding evaluation retains model, date, code revision, instructions, allowed tools, deadline and success rules. Separate execution failure, format validity, requirement satisfaction and human acceptance. The organization, model information and system configuration document different levels.
Worked example: shorter output does not lower the token rate
These are invented rates, not Claude prices. At two cost units per million input tokens and eight per million output tokens, 500 input and 1,000 output tokens cost (500×2 + 1,000×8) / 1,000,000 = 0.009. Keeping rates constant but reducing output to 600 tokens gives 0.0058, a reduction of (0.009 - 0.0058) / 0.009, approximately 35.6%.
This changes usage rather than token rates. Actual estimates require the relevant model, feature and delivery path in the pricing guide ↗, separating input, output, cache writes/reads, tools and retries.[5] This simple example excludes cache/tool costs. Check that shorter output retains required content. Token accounting differs from completion time.
Tool example: separate calls, authorization and completion
The tool guide ↗ distinguishes client tools executed by an application from server tools executed by Anthropic.[6] Similar names can hide different execution and result handling. Record the model's call, application permission check and actual completion separately.
Suppose a fictional draft-save tool times out after saving but before reporting its result. Blindly repeating the save can create duplicate drafts. Use a task identifier to inspect completion state, and test retries with mock tools. This example neither publishes anything nor asserts a product default. Read permissions and oversight alongside tool-assisted workflows.
Reconnect principles and evidence after model changes
A constitution describes intended principles, a system card evaluates a configuration and operational records document tests and observations. Distinguish them. Identify evidence to revisit after changes to models, instructions, material, tools or permissions. An earlier announcement's safety score does not automatically establish a new configuration's pass.
Separate regression replay from fresh-data testing; record failure discovery, risk judgments, legitimate completion and false blocking individually. Mark untested languages, files or tools unconfirmed. Track actual models and delivery/execution conditions rather than rely on an unchanged family name.[4][6]
SOURCES
Sources
-
Introducing ClaudeAnthropicOpen source ↗
-
Claude's ConstitutionAnthropicOpen source ↗
- Open source ↗
-
Models overviewAnthropicOpen source ↗
-
PricingAnthropicOpen source ↗
-
Tool use with ClaudeAnthropicOpen source ↗
KNOWLEDGE LINKS
Continue from here
BACKLINKS