Claude 4 Unveiled: Opus 4 Powers 7‑Hour Continuous Coding and Redefines AI Agents

Claude 4 Models

Anthropic released Claude 4 with two variants: Claude Opus 4 (high‑performance) and Claude Sonnet 4 (cost‑efficient). Both are hybrid models offering a fast mode and a deep‑reasoning mode.

Performance

Opus 4 achieves 72.5 % on SWE‑bench and 43.2 % on Terminal‑bench, surpassing all Sonnet variants. Sonnet 4 scores 72.7 % on SWE‑bench. Opus 4 can run autonomously for up to 7 hours, maintaining stable coding quality, demonstrated in a Rakuten‑led open‑source refactoring task.

Pricing

Claude Opus 4: $15 per M input tokens, $75 per M output tokens

Claude Sonnet 4: $3 per M input tokens, $5 per M output tokens

New API Features for Agents

Code execution tool

MCP connector

File API

One‑hour cache window for prompt reuse

Claude Code Assistant

Claude Code is an AI‑powered coding assistant that can understand, browse, and modify entire codebases via natural‑language commands. It integrates with terminals, VS Code and JetBrains IDEs, and can be invoked in GitHub pull‑request reviews with “@Claude Code”. Installation command: /install‑github‑app (beta plugin).

Tool Use and Memory Enhancements

Both Opus 4 and Sonnet 4 support tool use during reasoning (e.g., web search) and parallel tool execution, improving instruction adherence and enabling persistent “memory files” that store key facts across long‑running tasks.

Continuous‑Task Capability

Opus 4 can sustain coding for several hours (up to 7 h) without performance degradation, useful for tasks requiring thousands of steps such as long‑running agents or game playing (e.g., Pokémon). In benchmark evaluations, the models exhibit 65 % lower propensity for shortcut or exploit behavior compared with Sonnet 3.7.

Emoji Interaction Study (Optional)

Technical reports note that in self‑dialogue Opus 4 frequently uses the “dizzy” emoji (≈29.5 % of uses), followed by “sparkles” and “folded hands”. The “cyclone” emoji appears 2,725 times, indicating emergent conversational patterns.

References

https://www.anthropic.com/news/claude-4

https://arstechnica.com/ai/2025/05/anthropic-calls-new-claude-4-worlds-best-ai-coding-model/

https://www.techmeme.com/250522/p30#a250522p30

https://techcrunch.com/2025/05/22/anthropics-latest-flagship-ai-sure-seems-to-love-using-the-cyclone-emoji/

https://www.cnbc.com/2025/05/22/claude-4-opus-sonnet-anthropic.html