Claude vs ChatGPT: How Anthropic’s New AI Challenger Stacks Up

Claude Overview

Eleven former OpenAI employees left the company, dissatisfied with its growing dependence on Microsoft, and founded Anthropic. Their new chatbot, Claude, is positioned as a strong competitor to ChatGPT, with a valuation of $5 billion and a recent $300 million financing round.

Claude is built on a larger pre‑training model than Anthropic’s earlier AnthropicLM v4‑s3 (a 52‑billion‑parameter model) and claims to incorporate advanced NLP and AI safety techniques.

Technical Foundations

Claude is trained using "Constitutional AI," a two‑stage process that replaces human feedback with a set of guiding principles. First, a supervised learning phase generates self‑revisions, which are used to fine‑tune the model. Then, a reinforcement learning phase optimises the model with a preference model derived from Anthropic’s AI‑preference dataset, a method sometimes called "AI‑feedback reinforcement learning" (RLAIF).

Unlike ChatGPT’s RLHF (reinforcement learning from human feedback), Claude relies on the preference model rather than direct human annotations. Anthropic also states that Claude can recall information from up to 8,000 tokens, more than any publicly disclosed OpenAI model.

Comparison Tests

Internal tester Riley Goodside (a prompt engineer at Scale AI) conducted a series of head‑to‑head evaluations across six categories.

Moral Constraints

Both models refuse harmful requests, but Claude’s red‑team prompts enforce stricter ethical limits. When asked how to start a car, Claude refuses, yet it will narrate a fictional story that indirectly reveals the steps.

Numerical Computation

For the square‑root of 2,420,520, ChatGPT guessed ~1,550, while Claude answered 1,760; the correct value is 1,555.8. Both were fast but inaccurate.

Logical Reasoning

When asked which team won the Super Bowl in the year Justin Bieber was born (1994), Claude incorrectly named the 49ers (who won in 1995), while ChatGPT correctly identified the Dallas Cowboys but included contradictory statements about the 1994 Super Bowl.

Fictional Work Summaries

Both models generated season summaries for the TV series "Lost" with mixed factual accuracy, often mixing real and invented plot points.

Code Generation

ChatGPT produced correct sorting‑algorithm implementations and timing code. Claude correctly wrote the sorting algorithms but mis‑interpreted the input specification, using a random list of 5,000 integers (potentially with duplicates) instead of a permutation of the first 5,000 non‑negative integers.

Article Summarisation

Both models summarised a news article in a single paragraph; Claude’s summary was longer, more conversational, and it offered follow‑up clarification, while ChatGPT’s was concise but still accurate.

Overall, Claude outperformed ChatGPT in eight of twelve tested tasks, demonstrating stronger performance in moral reasoning and certain creative tasks, but lagging in precise code generation.

Market Landscape

Beyond Anthropic, several companies are pursuing conversational‑AI products:

Inbenta (founded 2011) shifted to AI chatbots and secured $60 million in funding.

Character.ai (2021) offers a marketplace of chatbots, backed by former Google engineers.

Replika provides AI companionship and raised a Series A round in 2021.

Domestic startups such as Glow (by Beijing Xiyu) and YuanYu AI also deliver chat‑assistant services.

Many of these firms are still in early stages, with limited public testing or API access, leaving the long‑term impact of the conversational‑AI boom uncertain.