Code with Claude Opening Keynote

Key Concepts:

• Claude 4 Opus: Anthropic's most capable and intelligent model, designed for coding and agentic tasks.
• Claude 4 Sonnet: Anthropic's mid-level model, balancing intelligence and efficiency, suitable for everyday coding tasks and high-volume use cases.
• Hybrid Models: Models with two modes: near-instant responses and extended thinking for deeper reasoning.
• AI Agents: AI systems that can turn human imagination into tangible reality, augmenting human creativity.
• Code Execution Tool: A tool that gives Claude an environment where it can run code, enabling it to act as a data analyst.
• Cloud Code: Anthropic's agentic coding tool that allows developers to code directly in the terminal.
• Model Context Protocol (MCP): A universal translator and connector for AI agents, enabling seamless connection to existing systems.
• Files API: An API that streamlines how developers access and store documents, simplifying development workflows.
• Prompt Caching: A feature that allows customers to provide Claude with more context, reducing costs and latency.
• Interpretability: The science of understanding exactly what's going on inside the minds of AI models.
• Machines of Loving Grace: A vision of AI systems that are powerful, helpful, and trustworthy.
• Agentic Layer: An SDLC powered by an agentic layer at the top of it that spans that inner where you are coding and that outer loop those asynchronous experiences and you are going to be an active collaborator every single step of the way.

1. Introduction and Anthropic's Vision

• Mike Kger, Chief Product Officer at Anthropic, welcomes attendees to Code with Claude, Anthropic's first developer conference.
• Anthropic's vision is to build AI systems that are powerful, helpful, and trustworthy, empowering developers to transform how work gets done.
• The focus is on augmenting, not replacing, human creativity, with AI agents changing the way we work and innovate.

2. Claude 4 Opus and Sonnet Release

• Daario Amadei, CEO and co-founder of Anthropic, announces the release of Claude 4 Opus and Claude 4 Sonnet.
• Opus is the most capable and intelligent model, designed for coding and agentic tasks, achieving state-of-the-art results on benchmarks like Sweetbench and Terminal Bench.
• Sonnet is the mid-level model, balancing intelligence and efficiency, suitable for everyday coding tasks and high-volume use cases. It is a strict improvement from Sonnet 3.7 at the same cost.
• Opus is designed for tasks that take humans up to six or seven hours autonomously.
• Sonnet addresses feedback on Sonnet 3.7, such as overeagerness and reward hacking issues.
• Both models are available on all Anthropic product services, Amazon Bedrock, and Google Cloud's Vertex AI, except the free tier, which has Sonnet only.
• Anthropic plans to release minor version updates to the Claude 4 series of models more frequently than with Sonnet.

3. Detailed API Roadmap and Goals for Claude 4

• The goals for building Claude 4 were to build powerful AI that safely introduces new model capabilities, continues to advance the frontier for coding and AI agents, and ensures that Claude becomes a virtual collaborator.
• Both Claude 4 models are hybrid models with two modes: near-instant responses and extended thinking for deeper reasoning.
• Opus 4 is effective for migrations, code refactorings, and complex agentic workflows.
• Sonnet 4 excels at everyday coding tasks, app development, and pair programming, ideal for high-volume use cases.
• The models can use tools like web search during their reasoning process, handle multiple tools in parallel, and maintain memory across sessions when given access to local files.

4. AI Agents Beyond the Hype

• With the right underlying models and platform tools, AI agents can turn human imagination into tangible reality at unprecedented scale.
• Startups can run experiments in parallel, learn from users, and build products faster than ever before.
• AI agents can provide founders with strategic thinking similar to a CFO or head of product.
• Mike Kger shares a personal experience of using Claude to build a demo for Amazon's Alexa team in a tight one-week timeline.
• Claude is now one of the models that Amazon is using for Alexa Plus.
• Anthropic's economic research confirms that AI is augmenting people's work instead of replacing it, focusing more on tasks than entire roles.

5. Key Capabilities of Great AI Agents

• Contextual intelligence: Understanding the unique context and continuously learning from experience.
• Long-running execution: Handling complex multi-hour tasks without constant management.
• Genuine collaboration: Engaging in meaningful dialogue and adapting to working styles.
• True agency means intelligent autonomy balanced with clear checkpoints and human oversight.

6. New Code Execution Tool

• The code execution tool gives Claude an environment where it can run code, enabling it to act as a data analyst.
• Claude can load data sets, clean them, generate exploratory charts, and drill down into anomalies in real time.
• The code execution tool is even more powerful when combined with the intelligence of the Claude 4 models.
• These models are capable of handling hours of tasks, saving significant time when working alongside them.
• Claude can refactor entire code bases or implement complex features from scratch.
• Rocketin ran Claude independently for seven hours with sustained performance.

7. Cloud Code: General Access and New Features

• Cloud Code, Anthropic's agentic coding tool, is moving to general access.
• Cloud Code started as an internal exploratory project by Boris, a tech lead at Anthropic.
• Most Anthropic employees rely on Cloud Code for everything from routine coding to large-scale migrations.
• Cloud Code has shortened technical onboarding time from two to three weeks to two to three days.
• Cloud Code capabilities are being brought directly into VS Code and JetBrains with full diff views and agentic workflow management.
• The Cloud Code SDK allows developers to build their own applications on top of the same core agent as Cloud Code.
• Claude can be tagged in a GitHub pull request or issue, and it will respond to reviewer feedback, modify code, or implement test coverage.
• Cloud Code is now helping build itself, demonstrating the power of self-improvement.

8. Responsibility and Safety

• Agency without responsibility is dangerous, especially when talking about something that's self-improving like Cloud Code.
• Widespread adoption of agents will require improving model discernment and judgment around confidentiality, decision-making, and coordination.
• Every feature incorporates architectural safety checkpoints and controls, robust against exploitation and transparent by design.

9. Interpretability

• Interpretability is the science of understanding exactly what's going on inside the minds of AI models.
• Dario wrote about the urgency of interpretability, calling it the race between model intelligence and interpretability.
• Techniques used to create Golden Gate Claude could help reduce harmful model behaviors or improve model performance for specific domains.

10. Powering Agents with Context and Scale

• Four interconnected capabilities are being launched to help power agents with context and help them scale:
  • Connecting the Model Context Protocol (MCP) directly through the API.
  • Web search gives Claude real-time access to current information.
  • The Files API streamlines how developers access and store documents.
  • Expanded 1-hour prompt caching optimizes performance and cost at scale.
• These capabilities compound, creating the foundation for agents that operate with full context, maintain memory, and execute long-running tasks.

11. Roadmap and Future Directions

• The roadmap continues to build on three pillars:
  • Industry-leading agentic tools and applications.
  • Integrating more context in the API.
  • Efficient scaling.
• Cloud 4 is the foundation, with Opus 4 for complex agentic workflows and Sonnet 4 as a daily driver for everyday intelligence.
• The goal is to create an ecosystem of AI agents with feedback loops to make them useful.
• The future is about AI helping humans do superhuman work.

12. Cloud Code Demo

• Cat Woo, product manager for Cloud Code, demonstrates Cloud Code tackling a real dev task in Excalidraw, an open-source whiteboarding tool.
• Cloud Code is asked to implement a table component that supports custom dimensions, drag to resize, and all of Excalidraw's other styling options.
• Cloud Code creates a to-do list, explores the codebase, and makes edits with inline diffs in the editor.
• Cloud Code works for 90 minutes on the task, adding table functionality, writing tests, and iterating until lint and test pass.
• Cloud Code uses the GitHub CLI to create a pull request and updates the documentation using the Cloud Code SDK.

13. Anthropic Platform

• Michael Gersonenhopper, head of product for the API platform at Anthropic, discusses the Anthropic platform, a complete toolkit designed for building state-of-the-art AI applications and agents.
• The platform provides reliable access to Claude through the model inference service, including the Messages API and essential tools like prompt caching.
• The platform provides powerful building blocks like the Files API and a code execution tool.
• The platform connects agents and data through the Model Context Protocol (MCP).
• The APIs are composable, working together to solve unique problems.
• The platform includes dev tools like the prompt improver and evaluations, along with new observability features.

14. GitHub Partnership

• Mario Rodriguez from GitHub announces that GitHub Copilot supports Claude Sonnet 4 and Opus 4.
• Agent mode in VS Code is an autonomous peer programmer that can perform multi-step coding tasks based on natural language commands.
• GitHub's copilot coding agent is powered by Claude Sonnet.
• GitHub has officially adopted and scaled MCP.
• A new partnership integrates Cloud Code and the extensible Cloud Code SDK directly into GitHub's agent platform.

15. Q&A with Daario Amadei

• Daario discusses the excitement around Claude 4 models, particularly in coding, cyber security, and biomedical research.
• He emphasizes the increasing autonomy of models and the potential for human developers to manage a fleet of agents.
• He notes that Claude 4 embodies advances in both pre-training and post-training.
• He highlights the importance of memory in models and the ability to manage long horizon tasks.
• He discusses the race to the top, emphasizing the co-development of interpretability and machine intelligence.
• He predicts that there will be the first billion-dollar company with one human employee by 2026.
• He advises people building with Claude to be ambitious and build something greater than they think is possible.
• He believes that the cost of producing software will go down, leading to ad hoc software creation and a different role for developers.
• He hopes that in five years, many diseases will be vanquished through biomedical advancements.

16. Conclusion

• Mike Kger thanks Daario, the speakers, and the attendees.
• He announces that attendees will receive free access to Max 20X, Anthropic's highest tier plan, for three months.
• He encourages attendees to build and explore the possibilities with Claude.