Anthropic Claude (1, 2)

Anthropic, founded in 2021 by ex-OpenAI researchers, launched Claude in March 2023 as a direct competitor to ChatGPT. Claude differentiated itself on alignment methodology (Constitutional AI in place of pure RLHF), honest refusals, long context (100K tokens at Claude 2's launch — unprecedented at the time), and a generally less-stiff conversational register. Claude is now a primary frontier model, with Claude 3, 3.5, 4, and beyond extending the line.

Claude 1 (March 2023)

Claude 1 was released as an API and via Slack integration. Capabilities:

• GPT-3.5–GPT-4 range depending on benchmark — generally competitive with GPT-4 on reasoning, behind on math and code.
• 9K–100K token context — much longer than GPT-3.5's 4K. Long-document QA was the headline use case.
• Honest refusals — when uncertain, Claude was trained to express uncertainty rather than confabulate. This reduced GPT-style hallucinations on the long tail.

Architecture details were not disclosed, following the post-GPT-4 closed-research norm.

Constitutional AI

The signature alignment methodology. Constitutional AI: Harmlessness from AI Feedback (Bai et al., Anthropic 2022) proposes replacing some of the human-labelled preference data in RLHF with AI-generated preferences guided by a written "constitution" — a list of natural-language principles like:

• "Choose the response that is most helpful, honest, and harmless."
• "Choose the response that is least likely to be discriminatory."
• "Choose the response that least encourages illegal activity."

The training pipeline:

1. SFT on human-written prompts and ideal responses.
2. AI revision — the model critiques and revises its own outputs against the constitution.
3. AI preference labelling — sample two responses, ask the model which better matches the constitution, train a reward model on these AI-generated preferences.
4. RL fine-tuning against the AI reward model.

The result: a model whose alignment is steered by a written document rather than implicit-in-the-data labellers' preferences. The constitution is auditable; the model's behaviour becomes more transparent. See Constitutional AI for the full mechanism.

Claude 2 (July 2023)

Claude 2 brought:

• 100K token context window — roughly 75K words. Could process whole books, codebases, long documents in one pass.
• Substantially better performance on bar exam, MMLU, GSM8K, HumanEval.
• Public consumer access at claude.ai (US/UK initially).
• Improved code generation and structured output.

The 100K context was the biggest practical differentiator at launch. Pre-Claude-2, frontier-LLM context was 4–32K. The jump to 100K opened new use cases — long-document analysis, large-codebase understanding — that ChatGPT could not handle until GPT-4 Turbo's 128K release in November 2023.

Anthropic's positioning

Anthropic explicitly positioned itself as a safety-focused frontier lab. Their "scaling-policy" documents (Responsible Scaling Policy, 2023) committed to capability evaluations and safety reviews tied to model deployment decisions. The framing — we will build frontier models, but with explicit safety protocols — became a common stance for frontier labs as 2023–24 progressed.

Practically, Claude's tone is more "thoughtful research assistant" than ChatGPT's "friendly product". Many academic and writing-heavy users prefer it; Claude's refusal style and honesty stance produce different downstream products.

What followed

• Claude 3 (March 2024) — three-tier release (Opus / Sonnet / Haiku), extended multi-modal vision input, ranked first on most public benchmarks at release.
• Claude 3.5 Sonnet / Opus (mid-2024) — substantial improvements at lower cost, particularly strong on coding and computer-use.
• Claude 4 (2025) — the next-generation frontier model, with continued emphasis on long-horizon agentic capabilities.

Claude is now one of three to four "frontier-tier" LLMs (alongside GPT-4o/o-series, Gemini, sometimes Grok), with Anthropic, OpenAI, and Google as the oligopolistic frontier-model providers.

What Claude established

• Constitutional AI as a credible alternative to pure RLHF.
• Long context as a flagship capability — 100K → 200K → 1M+ followed.
• The "thoughtful AI" register — distinct from ChatGPT's voice, often preferred for analytical work.
• Anthropic as a serious frontier competitor, ending OpenAI's monopoly on top-tier chatbots in 2023.

What to read next

• Constitutional AI — the alignment methodology in detail.
• GPT-4 — the contemporary OpenAI frontier model.
• Claude 3 — the successor.