Claude 1: How Anthropic's First Model Was Actually Built

A Different Company, a Different Default

This site's Model Case Study series has so far traced one lab's trajectory from full openness to near-total opacity. Anthropic's story starts from the opposite default: Claude was closed-architecture from its very first public release on March 14, 2023 — no disclosed parameter count, no released weights, ever, for any Claude model. Understanding why requires starting not with the model, but with why the company exists at all.

Founded Specifically to Disagree About Pace

Anthropic was founded in 2021 by Dario Amodei (previously OpenAI's VP of Research) and Daniela Amodei (previously OpenAI's VP of Safety & Policy), alongside a group of other former OpenAI researchers. Their stated reasoning was consistent across public accounts: they believed capability was scaling faster than the field's ability to understand and control it safely, and wanted to build a company where safety research was the founding purpose, not an add-on. This is a fundamentally different starting problem than any article in this site's GPT lineage — Anthropic's "research trigger" for existing at all was itself a safety disagreement, before any specific model was ever discussed.

Constitutional AI: Less Human Labeling, More Written Principles

Anthropic's foundational research bet, published as "Constitutional AI: Harmlessness from AI Feedback" (Bai et al., December 2022) — just three months before Claude's public launch — proposed training a model to critique and revise its own outputs against a written set of principles (a "constitution"), then using that self-critique to generate preference data for reinforcement learning, rather than relying entirely on human labelers to judge every harmful-vs-harmless comparison. This built directly on Anthropic's earlier "Training a Helpful and Harmless Assistant with RLHF" (Bai et al., April 2022), which established the human-feedback methodology Constitutional AI then partially automated.

Eight People Who Walked Away From OpenAI

Anthropic's founding team included Dario Amodei (CEO), Daniela Amodei (President), Benjamin Mann, Jared Kaplan, Jack Clark, Sam McCandlish, Tom Brown, and Chris Olah — the last three names are worth pausing on directly. Tom Brown was the lead author of the GPT-3 paper this site's own case study covers in depth; Sam McCandlish and Jared Kaplan were co-authors on foundational OpenAI scaling-law research; and Chris Olah is one of the field's most prominent interpretability researchers, a research direction this site's Interpretability article covers. This is a genuinely unusual founding story within this entire case-study project: several of the people who built GPT-3's scale-driven playbook left to found a company explicitly built around questioning how fast that playbook should be run.

Closed From the First Release, Not Eventually

Claude launched on March 14, 2023 via the Anthropic API in two variants: the full Claude model and a faster, cheaper Claude Instant model, both with a 9,000-token context window at launch (expanded to 100,000 tokens on May 11, 2023). Unlike GPT-1, GPT-2, or GPT-3 — each of which disclosed parameter counts and, for GPT-1 and GPT-3, detailed architecture — Anthropic has never publicly disclosed a parameter count, layer architecture, or training compute figure for any Claude model, Claude 1 included. This is the single most important structural difference between the two labs' case-study series: GPT's opacity was a trend that developed over time (transparent through GPT-3, closed from GPT-4 on); Claude's opacity was the starting condition.

RLHF, Then Constitutional AI on Top

While Claude 1's specific training dataset was never disclosed, Anthropic's published methodology papers describe the general training technique: a base language model pretrained on a large text corpus, fine-tuned via reinforcement learning from human feedback for helpfulness, with Constitutional AI's self-critique process specifically targeting the harmlessness side of that tuning — reducing how much of the harm-avoidance training required direct human judgment on potentially disturbing content. This connects directly to this site's On-Call Reality article's underlying point about protecting the humans doing safety-critical work: fewer human labelers exposed to extreme harmful-content examples is a real, if rarely discussed, working-conditions benefit of the Constitutional AI approach, not just a training-efficiency one.

Before the Responsible Scaling Policy Existed

Claude 1 predates Anthropic's formal Responsible Scaling Policy (RSP) and AI Safety Level (ASL) classification system, both introduced later in 2023 — this series' forthcoming Claude 4 case study covers the first ASL-3 activation in detail. At launch, Claude 1's safety approach was Constitutional AI itself: the harmlessness training technique functioned as the safety framework, rather than a separate, named risk-classification system layered on top of it. This is worth naming honestly as a real limitation of treating Claude 1 through the same "safety framework" lens this series applies to later models — the framework, in this case, hadn't been formally named yet.

API-First, Partner-Distributed

Claude 1 launched directly as a closed API product, not as a research paper with a public demo — distributed initially through early partners rather than as a standalone consumer product. This is a genuinely different release shape than GPT-1's paper-plus-GitHub-release, and closer in spirit to how this site's Compute Economics article describes commercial API access becoming the default distribution model for capable models — except Anthropic started there from day one, rather than arriving at it after three more open generations the way OpenAI's GPT lineage did.

The Technique Outlived the Model

Claude 1 (versions 1.0 through 1.3) and Claude Instant 1.x were deprecated on September 4, 2024 and fully retired on November 6, 2024 — a relatively short public lifespan. What carried forward wasn't the specific model, but Constitutional AI as a technique, and the closed-architecture-from-day-one norm that every subsequent Claude release, through Claude 2, the Claude 3 family, and beyond, has maintained without exception.

Readiness Checklist

⚠️ What's Missing or Uncertain

Where This Case Study Goes Next

Claude 2 arrives four months later, in July 2023, expanding the context window further and marking Anthropic's first real push toward a broader consumer and enterprise product, rather than a partner-distributed API alone.

🔗 Reference Links

🎥 Recommended Videos

🧭 Closing — A Company Built Around a Disagreement