Anthropic Releases Claude 4: How Far Can an AI Focused on 'Safety' Go?

Anthropic has released Claude 4, directly competing with GPT-4o.

Capability Improvements

• Reasoning: Ties with GPT-4o on benchmarks such as GPQA and MATH
• Coding: Scores above GPT-4o on HumanEval for the first time
• Long context: 200K tokens (can process an entire book at once)
• Safety: 30% higher refusal rate than GPT-4o in red-team testing

Business Model

Anthropic's strategy differs from OpenAI's: - Focuses on the enterprise market (industries with high compliance requirements, such as finance, law, and healthcare) - Deeply integrated with AWS (Amazon invested $4 billion) - "Responsible AI" as a brand label

Challenges

Claude's market share among developers remains relatively low. Most developers are still accustomed to using ChatGPT or the GPT API. Anthropic needs to find differentiated use cases to break through.

That said, their safety research is indeed solid, and Dario Amodei (CEO) is one of the most important voices in the AI safety field. With AI regulation becoming increasingly stringent, Anthropic's "safety card" could be a long-term advantage.

References:

• Anthropic — Introducing Claude 4 (2025-03)