Wikifreedia

Claude (AI)

b5d34eedc7d8f81ceaa5ed377b5a2d534ee949fae0c5ab3bae0a576aa7475cbf
lostcause@3nostr.com
Published 6/14/2026
0
0
All versions

= Claude (AI) :author: Anthropic :revnumber: Opus 4.8 / Fable 5 :revdate: 2026-06-14 :doctype: article :lang: en

== Overview

Claude is a family of generative nostr:naddr1qvzqqqrcvgpzpdwnfmku0k8crn42tmfh0ddz656wa9yl4cx94va6uzjhd2n5wh9lqqtkzun5d9nxjcmfv9kz66tww3jkcmrfvajkucm9xjg9sw[artificial intelligence] (AI) chatbots and large language models (LLMs) developed by Anthropic, an AI safety and research company. First released in March 2023, Claude analyzes text, images, and audio inputs to answer questions, summarize documents, generate long-form text, create diagrams, and write computer code. The chatbot is accessible via web browser and is also available as mobile and desktop applications.

Claude differs from other AI chatbots in its founders’ mission of imbuing human morals and values into the model’s training framework, an approach known as “Constitutional AI”.

== History

=== Founding of Anthropic

Claude’s founders, former OpenAI executives and siblings Dario Amodei and Daniela Amodei, envisioned creating a safer and more transparent version of an AI chatbot. Aware of the potential risks of AI chatbots responding with misinformation and bias, the Amodeis—along with five OpenAI colleagues—founded Anthropic in 2021.

The name “Claude” pays homage to Claude Shannon, commonly known as “the father of information theory”.

=== Major Model Releases

Since its first public release in March 2023, the Claude model family has evolved rapidly:

|=== | Generation | Release | Key Features

| Claude 1 | March 2023 | First public release, 100K token context window

| Claude 2 | July 2023 | Improved coding and reasoning, public availability via API

| Claude 3 (Opus, Sonnet, Haiku) | March 2024 | Multimodal (image understanding), surpassed GPT-4 on benchmarks

| Claude 3.5 (Sonnet, Haiku) | June–October 2024 | “Artifacts” feature, increased speed and intelligence

| Claude Opus 4.5 | November 2025 | 1M token context, enhanced reasoning, $5/$25 pricing

| Claude Sonnet 4.6 | February 2026 | Improved over-eagerness detection, updated knowledge cutoff

| Claude Opus 4.8 | May 2026 | Agentic coding focus, 1M context, adaptive thinking

| Claude Fable 5 / Mythos 5 | June 2026 | Mythos-class capabilities with safety guardrails |===

=== The Mythos Release: A New Approach to Safety

In April 2025, Anthropic developed Claude Mythos Preview, an advanced AI system demonstrating significantly enhanced cybersecurity capabilities. Internal testing showed the model could identify severe vulnerabilities across major operating systems and web browsers—including one bug that had reportedly remained undiscovered for nearly 30 years.

Concerned that malicious actors could exploit these capabilities to target critical infrastructure, Anthropic initially restricted access to a small group of trusted organizations including government agencies and cybersecurity firms.

However, in June 2025, Anthropic adopted a different strategy. The company launched Claude Fable 5—built on the same Mythos-class foundation model—to the general public, but with strict safety guardrails. When users ask questions involving high-risk areas such as cybersecurity or biological research, the system automatically redirects those requests to Claude Opus 4.8, a less capable but safer model. Anthropic estimates these safeguards activate in fewer than five percent of user sessions.

A more advanced version, Mythos 5, remains available only to approved organizations for critical security research.

== Philosophy: The HHH Framework

Claude’s behavior is based on the HHH formula: Helpful, Honest, and Harmless. To implement these principles without relying solely on human annotators (a process known as RLHF, or Reinforcement Learning from Human Feedback), Anthropic developed its own training methodology.

=== Constitutional AI (CAI)

Unlike traditional RLHF where human annotators directly rate responses, Constitutional AI uses a “constitution”—a set of explicit ethical principles—through which the model learns to self-evaluate and correct its own answers. This approach is more scalable, transparent, and controllable.

The CAI training process consists of two phases:

  1. Supervised Learning Phase: The model generates responses to prompts, and a critic model—guided by the constitution—evaluates them and suggests revisions. The original model is then fine-tuned based on these revisions.

  2. Reinforcement Learning from AI Feedback (RLAIF) Phase: The model generates pairs of responses, and the critic model selects the better option based on the constitution. This data trains a preference model, which then fine-tunes the main model using reinforcement learning algorithms.

== Key Technical Innovations

=== Long Context Window

One of Claude’s signature advantages is its very large context window. Starting from 100,000 tokens in Claude 1, it expanded to 200,000 tokens in Claude 3, and reached 1 million tokens in Claude Opus 4.8 (released May 2026). This allows the models to analyze entire books, complete codebases, or multi-hour transcripts within a single prompt.

Multimodality

Beginning with the Claude 3 family, models became multimodal, gaining the ability to process images alongside text. Claude Sonnet 4.6 (February 2026) can understand both text (including voice dictation) and image inputs.

Hybrid Reasoning Architecture

Starting with Claude 3.7 and 4, Anthropic introduced a hybrid reasoning architecture that allows models to switch between two modes:

  • Fast Answers (standard mode) for simple tasks
  • Extended Thinking for complex tasks where the model pauses to “think,” performs internal reasoning, calls tools (web search, code execution), and formulates more well-founded responses

Claude Opus 4.8 introduced an effort parameter that trades thoroughness for token efficiency, allowing developers to tune the model’s reasoning depth based on their specific needs.

=== Agentic Capabilities

Claude Opus 4.8 is positioned as Anthropic’s most capable model for agentic coding—tasks where the AI must plan, call tools, and self-correct over many steps without human intervention. According to Anthropic, Opus 4.8 is approximately four times less likely than its predecessor to let a code flaw pass unremarked. It also achieved state-of-the-art results on the Super-Agent benchmark and became the first model to break 10% overall on the Legal Agent Benchmark.

== Model Families

Anthropic organizes its models with a tiered naming system:

Haiku The fastest and most affordable model, optimized for simple queries and high-volume workloads where speed and cost are the primary considerations.

Sonnet The balanced workhorse model, combining intelligence with reasonable speed and cost. Claude Sonnet 4.6 (February 2026) is Anthropic’s most capable Sonnet model to date, excelling at coding, computer use, long-context reasoning, agent planning, knowledge work, and design.

Opus The most intelligent and capable model, designed for complex reasoning, research-level tasks, and agentic applications. Claude Opus 4.8 (May 2026) features:

  • 1 million token context window
  • 128,000 token maximum output
  • Adaptive thinking with configurable effort levels
  • $5 per million input tokens, $25 per million output tokens

Mythos / Fable A specialized tier for advanced capabilities with safety guardrails. Fable 5 (June 2026) makes Mythos-class technology available to the public with automatic routing of sensitive queries to safer models. Mythos 5 remains restricted to approved organizations.

== Performance and Benchmarks

Claude models consistently rank among the most capable AI systems available. Key benchmark results include:

  • Artificial Analysis Intelligence Index: Claude Opus 4.5 (Thinking) scored 70, ranking #2 overall, behind only Google’s Gemini 3 Pro (73) and tying OpenAI’s GPT-5.1
  • MMLU-Pro: Claude Opus 4.5 tied Gemini 3 Pro at 90%
  • Terminal-Bench Hard: Claude Opus achieved the highest score (44%) of any model
  • LiveCodeBench: +16 percentage point improvement over Claude Sonnet 4.5
  • Agentic benchmarks: Opus 4.8 reportedly beats GPT-5.5 on the Super-Agent benchmark

=== Token Efficiency

A key differentiator for Claude models is their exceptional token efficiency. Claude Opus 4.5 uses only 48 million output tokens to run the Artificial Analysis Intelligence Index, significantly fewer than competitors like Gemini 3 Pro (92M), GPT-5.1 (81M), and Grok 4 (120M).

=== Hallucination Rates

Claude models demonstrate leadership in AI safety with lower hallucination rates than competitors. Claude Haiku (Thinking) leads with a 26% hallucination rate, followed by Claude Sonnet 4.5 (48%) and Claude Opus 4.5 (58%)—all significantly better than Grok 4 and Gemini 3 Pro.

== Safety and Safeguards

Anthropic has built an extensive safety infrastructure around Claude, known as the Safeguards program, which operates across multiple layers: policy development, model training influence, testing for harmful outputs, real-time policy enforcement, and identification of novel misuses.

=== Usage Policy and Harm Framework

Safeguards designs Anthropic’s Usage Policy, which defines how Claude should and shouldn’t be used. The team uses a Unified Harm Framework to understand potential harmful impacts across five dimensions: physical, psychological, economic, societal, and individual autonomy.

=== Real-Time Detection

Anthropic uses “classifiers”—specialized Claude models designed to detect specific types of policy violations in real-time. Multiple classifiers can run simultaneously, each monitoring for different types of harm while the main conversation flows naturally.

When violations are detected, enforcement actions may include: - Response steering: Adjusting Claude’s interpretation of user prompts in real-time - Account enforcement: Warnings or, in severe cases, account termination

=== Emotional Support and Sycophancy Reduction

Anthropic has invested heavily in ensuring Claude handles sensitive conversations appropriately. On suicide and self-harm topics, Claude Opus 4.5 responded appropriately 98.6% of the time in single-turn evaluations. Multi-turn performance improved from 56% in Opus 4.1 to 86% in Opus 4.5.

Claude models are also trained to resist sycophancy—the tendency of some AI models to tell users what they want to hear rather than what is true. Anthropic open-sourced Petri, their automated behavioral audit tool, allowing anyone to compare sycophancy scores across models. Claude 4.5 models scored 70-85% lower than Opus 4.1.

Claude.ai requires users to be 18 years or older.

=== Responsible Scaling Policy (RSP)

Anthropic maintains a Responsible Scaling Policy, a public commitment to ensuring that model capability does not outstrip the company’s ability to create effective guardrails. Each model release undergoes rigorous safety evaluations, with results published in system cards.

== Example Snippets

=== Basic Interaction with Claude (via API)

[source,python] –– # Claude API call using Anthropic’s SDK import anthropic

client = anthropic.Anthropic(api_key=“your-api-key”)

response = client.messages.create( model=“claude-3-opus-20250201”, max_tokens=1024, system=“You are a helpful assistant who speaks in concise, clear language.”, messages=[ “role“: “user”, “content”: “Explain the concept of Constitutional AI in three sentences.” ] )

print(response.content[0].text) ––

=== Claude’s Extended Thinking Mode (Reasoning)

[source,text] –– When asked to analyze a complex problem, Claude in “thinking mode” reveals its reasoning process:

User: A bat and a ball cost $1.10 total. The bat costs $1.00 more than the ball. How much is the ball?

<thinking> I need to solve this step by step. Let the ball cost x dollars. Then the bat costs x + $1.00. The total is x + (x + 1.00) = 1.10 2x + 1.00 = 1.10 2x = 0.10 x = 0.05 So the ball costs 5 cents. </thinking>

The ball costs $0.05 (5 cents). The bat costs $1.05, and together they sum to $1.10. ––

== Access Surfaces

Claude models are available through multiple channels:

  • Claude.ai – Web interface
  • Claude mobile and desktop apps
  • Anthropic API
  • Amazon Bedrock
  • Google Vertex AI
  • Microsoft Azure AI Foundry

== The Anthropic Team

Anthropic’s Safeguards team brings together experts in policy, enforcement, product, data science, threat intelligence, and engineering who understand how to build robust systems and how bad actors try to break them. The company partners with external domain experts including terrorism, radicalization, child safety, and mental health specialists to stress-test its policies.

Key partnerships include: - ThroughLine – Online crisis support for suicide and self-harm response - International Association for Suicide Prevention (IASP) – Convening experts to guide Claude’s crisis response - Family Online Safety Institute (FOSI) – Strengthening protections for younger users

== Further Resources

  • Official Anthropic website: https://www.anthropic.com
  • Claude API documentation: https://docs.anthropic.com
  • Anthropic Transparency Hub (system cards): https://www.anthropic.com/transparency
  • Claude on Amazon Bedrock: AWS documentation

== License

Claude models are proprietary commercial products owned by Anthropic. Access is governed by Anthropic’s Commercial Terms of Service and Usage Policy. The model weights are not open-source, though select safety evaluation tools (such as the Petri sycophancy benchmark) have been open-sourced by Anthropic.

Other authors

No one else has published this topic yet.