Claude Opus 4.5: The Engineers' Model

Detailed Summary

The King is Back (0:00 - 0:28)

The video introduces Claude Opus 4.5 as a groundbreaking AI model for engineers, positioning it as a superior tool for agentic engineering. It highlights the model's immediate impact, suggesting it surpasses other models like Gemini 3 due to its specialization in engineering delegation and autonomous tasks.

Claude Opus 4.5 Capabilities (0:28 - 5:03)

• Opus 4.5 offers two key advantages: enhanced agent delegation and the ability to run longer, harder, more complex tasks.
• The model is highly effective at managing a team of sub-agents, enabling well-coordinated multi-agent systems.
• A demonstration shows an Opus 4.5 agent kicking off five Opus 4.5 sub-agents simultaneously, each operating its own browser to perform tasks like system card analysis and model overview checks.
• The workflow involves a primary agent prompting sub-agents, which then respond back to the primary agent, and finally, the primary agent responds to the user.
• Anthropic is training Opus 4.5 to be a better prompt engineer for sub-agents, emphasizing the importance of better agents, then more agents.
• The model's specialization in engineering tasks allows for scaling compute to scale impact, with agents calling other agents being a significant future trend.

Incredible Model Pricing (5:03 - 7:04)

• Anthropic has slashed the pricing for Opus 4.5 by a third, making it a state-of-the-art model with premium pricing for premium compute.
• The previous pricing for Opus 4.1 was $15 per million input tokens and $75 per million output tokens.
• Despite being a premium model, Open Router reports Opus at approximately 60 tokens per second, indicating it is both cheap, fast, and state-of-the-art for engineering work.
• The presenter emphasizes that valuable things are not free, and the reduced cost combined with high performance makes Opus 4.5 the best model for software engineering work.

Opus 4.5 Browser Automation Results (7:04 - 11:23)

• The video demonstrates browser UI testing results, where the primary agent, using a markdown file as a prompt, delegates tasks to sub-agents.
• Tasks included summarizing releases, screenshotting images, getting pricing information, and summarizing orchestration details from the Opus 4.5 system card.
• The agents successfully downloaded and named all images from the webpage, showcasing detailed delegation capabilities.
• Another sub-agent looked up and summarized pricing for current generation Claude models, explaining their optimal use cases (e.g., Haiku for speed/cost, Opus as the new workhorse).
• The browser automation tasks are run within the agent sandbox skill codebase, which has been enhanced for better agent operation without blowing away the context window.

On the fly SUBAGENT browser testing (11:23 - 21:32)

• Agentic browser testing at scale allows Opus 4.5 to accomplish complex engineering work repeatedly.
• Testing is crucial for increasing review velocity, one of the primary constraints in agentic coding (planning and reviewing).
• Opus 4.5 can one-shot full-stack applications, as demonstrated by building five sandboxes, each hosting a different application.
• The workflow for building these applications involves planning, building, hosting, and then having agents test the hosted application, with a feedback loop for improvements.
• Examples of applications built by Opus 4.5 include a voice notes application using 11 Labs' Scribe 2.5, a graphing tool, and a decision-making matrix tool.
• These applications run in dedicated agent environments (sandboxes, using E2B as a host), providing isolation, scale, and autonomy.
• The presenter highlights that a single complex prompt can be used to build and modify full-stack applications, emphasizing the evolution of engineering work.

Agentic Prompt Breakdown (21:32 - 32:10)

• The video details a generic browser test prompt that uses N user story workflows, where sub-agents execute each workflow in parallel.
• The prompt format emphasizes consistency, with clear instructions and a detailed workflow for the agents to follow.
• The demonstration shows multiple agents operating browsers in headed mode, performing specific user stories and validating work on the application UI.
• This dynamic, natural language interface for testing is presented as a significant advancement over traditional deterministic testing frameworks.
• The core idea is that engineering is changing from what an individual can do to what they can teach their agents to do, with agents becoming the new compositional unit.
• The progression for mastering agents involves learning to operate a single agent, then a better agent (through prompt and context engineering), then more agents (sub-agents, custom agents), and finally, orchestration.
• Agent sandboxes provide isolation, scale, and autonomy, enabling powerful language models to perform complex tasks.
• The presenter emphasizes that Opus 4.5 allows for delegating more work and running longer, more challenging workflows than ever before.
• The video concludes by advocating for building systems that build systems, focusing on agents that run applications rather than building applications directly.

Notable Quotes

"Engineers, the king is back. This model is like that top tier engineer. When they walk into the meeting, everyone shuts up." — [Presenter]

"Opus 4.5 is also very effective at managing a team of sub agents enabling the construction of well-coordinated multi-agent systems." — [Presenter]

"Valuable things are by nature not free. If something is free and it is valuable, someone put a lot of work into making it that way for you or you are the product." — [Presenter]

"Anthropic is putting their foot down and they're saying the most important thing for these models is to be great at engineering." — [Presenter]

"If you master the agent, you master knowledge work. If you master the agent, you master engineering." — [Presenter]

"Scale your compute to scale your impact." — [Presenter]

"You want to build the system that builds the system, right? Build the agents that run your application. Don't build the application yourself anymore. You have agents for that." — [Presenter]

"We can delegate more and better than ever. And we can run longer, more challenging, more complex workflow with powerful models like Gemini 3 and especially the new Claude Opus 4.5 running in Claude Code." — [Presenter]

"It's not just about model intelligence anymore. All right. It's about the agent harness. So you know, cloud code. What are you putting the model in? And it's about the agentic tooling that you give your agents to operate." — [Presenter]

"One agent is not enough." — [Presenter]

"This is for mid to senior level engineers. If you're a new engineer or if you're a vibe coder, this is not for you. You have to be a cracked vibe coder to understand what's happening here." — [Presenter]

"What can we do with this technology that we couldn't do before?" — [Presenter]

"Stay focused and keep building." — [Presenter]