Big news! Good news for developers, the coding king Claude 4 is here!

Just now, Anthropic held their first developer conference "Code with Claude". The company's Chief Product Officer Mike Krieger (who previously worked on Instagram) and CEO and co-founder Dario Amodei both came out to support the event.

Krieger said at the beginning that Anthropic's goal is to create AI systems that are powerful, helpful, and reliable, and to allow developers to use these good things to improve their working methods and even the way they start a company. The focus is on enhancing human creativity, not replacing humans.

At this launch event, Anthropic launched the new generation of Claude 4 series models, including Claude 4 Opus and Claude 4 Sonnet, which will be available on all related products on the day of launch.

01 Claude 4 Opus

Positioned as the most capable and intelligent model in the series

Claude 4 Opus is particularly suitable for encoding and handling complex agent tasks. The model performs at the top of some industry standard tests (such as Sweep Bench) and can autonomously complete tasks that previously took humans six to seven hours to complete. Amodei mentioned that he was confused by the realism of the internal minutes generated by Opus and mistakenly thought they were written by team members.

02 Claude 4 Sonnet

A mid-range model that strikes a balance between intelligence and efficiency

Claude 4 Sonnet is a direct upgrade from Sonnet 3.7, offering greater intelligence at the same cost. The new version also addresses feedback that Sonnet 3.7 was "overzealous" (performing tasks beyond what was required).

Both Claude 4 models are "hybrid models" that support both fast response and extended thinking for deep reasoning. They are available on platforms such as Claude's own products, Claude Code tools, Anthropic API, Amazon Bedrock and Google Cloud Vertex AI.

03 AI Agent

Anthropic believes that AI agents are changing the way we work and innovate. The new Claude 4 model brings several key new features:

• Tool usage and parallel processing: The model can use tools such as web search during inference and can process multiple tools in parallel.

• Cross-session memory : After gaining access to local files, the model can maintain memory across sessions and gradually accumulate knowledge.

• Long-term task execution capability: Compared with the minute-level autonomous working capability of Claude 3 and the approximately 45 minutes of autonomous working capability of Claude 3.7, Claude 4 can achieve autonomous task processing for up to several hours, and some customers reported that it can work continuously for 7 hours.

At the same time, Chief Product Officer Mike Krieger pointed out that an excellent AI agent should have three core capabilities:

1. Contextual Intelligence: Understand the unique context of users and organizations and continuously learn from experience. He emphasized that the 100th task of interacting with an intelligent agent should be far better than the first task.

2. Long-running Execution: Process complex multi-hour tasks without constant management and coordinate with other agents and humans as needed.

3. Genuine Collaboration: Engage in meaningful conversations, adapt to user work styles, and provide transparent reasoning for their actions.

04 Developer Tools and API Upgrades

Anthropic has launched a series of new API features and tool updates for developers:

• Code Execution Tool: Available on the Anthropic API, this tool enables Claude to run code and play the role of a data analyst, turning raw data into visual insights.

• Claude Code official version and expansion: This AI programming tool, originally initiated by Boris, the internal technical director of Anthropic, has now entered the official access stage . It is said that this tool has shortened the technical learning time of Anthropic’s internal engineers from two or three weeks to two or three days.

• IDE integration: Supports common editors such as VS Code and JetBrains, and code changes can be viewed directly inline in the editor.

• Claude Code SDK: provides a development kit that allows developers to integrate Claude Code's features into their own applications and workflows. For example, Anthropic demonstrated how to use a tag (@Claude) on GitHub to let Claude Code respond to code review (PR) feedback, modify code, or implement test coverage.

At the press conference, the use of Claude Code to add a table component to the open source whiteboard tool Excalidraw was demonstrated. It was said that Claude Code completed the task autonomously in 90 minutes.

API platform upgrade:

1. Files API: Available today in the API, it allows developers to access and store documents, and provides a cookbook to help developers build memory capabilities into their apps.

2. Model Context Protocol (MCP) : Now you can connect directly via API. The protocol has been adopted by companies such as Microsoft, Google, OpenAI, and the community has built more than 3,000 integrations.

3. Web Search: Provides Claude with real-time information access capabilities.

4. Prompt Caching upgrade: The cache lifespan (TTL) is extended from 5 minutes to 1 hour. For long prompt tasks, this can reduce costs by up to 90% and latency by 85%. It is said that more than 50% of the input tokens on the platform have been cached.

05 Deepening cooperation with GitHub

1. GitHub Copilot now supports Claude Sonnet 4 and Opus 4 and is now available.

2. The provider coding agent within GitHub is driven by Claude Sonnet, who was chosen for his strong coding knowledge, problem-solving skills, instruction-following abilities, and caching support for the Anthropic API.

3. GitHub is officially adopting and promoting the MCP protocol.

4. Claude Code and its SDK will be integrated into GitHub's agent platform.

Last but not least

Anthropic's "Code with Claude" conference has a lot of valuable information, especially coding, which is good for developers.

At the same time, not only did they release a new model, but more importantly, they showed how they want to help developers make good use of AI agents.

They emphasize that AI agents must be able to work for a long time, understand specific situations, and collaborate well with people. They also focus on safety and explainability, which shows that they really want to make AI a "collaborator" that can be of great use.

This is not only about building tools, but also about shaping a new and smarter production relationship, heralding a new starting point for developers and AI to create a future together.

2025, Agent!