Anthropic has announced the release of two new AI models, Claude Opus 4 and Claude Sonnet 4, designed to enhance coding, reasoning, and AI agent performance.

The models aim to improve tasks that require extended reasoning and tool usage.

According to the company, these models feature improved memory capabilities and enhanced instruction-following.

Both models are accessible through the Anthropic API, Amazon Bedrock, and Google Cloud’s Vertex AI.

Pricing remains consistent with previous models, with Opus 4 priced at US$15/US$75 per million tokens and Sonnet 4 at US$3/US$15.

🔗 Source: Anthropic

🧠 Food for thought

1️⃣ AI coding models are evolving beyond assistants to become collaborative partners

The shift in language from “AI coding assistants” to models that can sustain “long-running tasks” and “work continuously for several hours” reflects a fundamental evolution in how AI participates in development workflows.

According to Stanford’s 2025 AI Index Report, AI has moved from simple code completion to executing complex development tasks autonomously, transforming how software is created 1.

This progression mirrors broader industry trends, with predictions that 70% of new applications will be developed with AI assistance by 2025, as developers transition from code producers to strategic architects who leverage AI capabilities 2.

Multiple partners cited in the announcement—including Cursor, Replit, and Block—specifically highlight Claude Opus 4’s ability to understand complex codebases and maintain performance over extended periods, suggesting a shift toward true collaborative intelligence rather than just supplementary assistance.

The introduction of “extended thinking with tool use” represents a significant advancement where Claude can alternate between reasoning and accessing external resources, reflecting how human developers work through challenging problems.

2️⃣ Standardized benchmarks have become crucial differentiators in the AI coding race

Anthropic’s emphasis on Claude Opus 4’s performance on SWE-bench (72.5%) and Terminal-bench (43.2%) reflects the growing importance of standardized metrics in evaluating AI coding capabilities.

The AI coding landscape has become increasingly competitive, with multiple comprehensive comparisons published in 2025 evaluating between 10-15 different models on specific coding tasks and benchmarks 3 4.

Recent Anthropic developments