Skip to content
AnthropicAnthropic

Claude Code updates: When to use Haiku 4.5, Claude Code on web, and more.

We're covering everything new with Claude for developers, including the launch of Claude Code for the Web, Claude Haiku 4.5, and more. Join the Claude Developer Discord - https://anthropic.com/discord Try Claude Code on the Web - https://claude.ai/code 00:02 Claude Code on the Web 00:42 Claude Code on the Web demo 01:26 Claude Haiku 4.5 02:10 When to use Haiku 4.5 02:51 Claude Code on the Web follow-up 03:05 Claude Code on the Web teleport 04:44 Summary

Oct 30, 20255mWatch on YouTube ↗

CHAPTERS

  1. Claude Code on the web: browser- and mobile-first coding sessions

    Claude Code is now accessible from the web and the Claude mobile app, so you don’t need to open a terminal or IDE to delegate coding work. You connect GitHub, pick a repository, describe the task, and Claude executes it on Anthropic’s cloud infrastructure.

    • Available at claude.ai/code and via Claude mobile app
    • Connect GitHub and select a repository to start
    • Describe the work in natural language; Claude executes in the cloud
    • Designed for quick delegation without opening local dev tools
  2. Parallel cloud development across multiple repositories

    A core benefit of Claude Code on the web is running multiple tasks simultaneously, even across different repos. This is positioned as a way to chew through bug backlogs, routine fixes, or new feature ideas more efficiently.

    • Run multiple Claude Code sessions in parallel
    • Work can span different repositories at once
    • Useful for bug backlogs, routine maintenance, and new features
    • Cloud execution means local machine constraints are reduced
  3. Web demo: adding registration input validation and strong password checks

    The presenter demonstrates a concrete task: adding input validation to a user registration endpoint, including strong password validation on both frontend and backend. The session kicks off from the web UI with a single prompt.

    • Example prompt targets both backend endpoint and frontend validation
    • Task initiated directly from the browser after selecting a repo
    • Emphasis on end-to-end changes across stack (API + UI)
    • Shows “prompt → execution” workflow for typical app changes
  4. How Claude executes: isolated sandbox with full Git access and live progress

    As the task runs, the video explains that Claude operates in an isolated sandbox environment with full Git access. Users can monitor the work as it progresses in real time.

    • Claude runs in an isolated sandbox environment
    • Full Git access enables real repo modifications
    • Progress is visible live while the agent works
    • Supports an “agentic” workflow rather than single-shot completions
  5. Model update: introducing Claude Haiku 4.5 after Sonnet 4.5

    The video pivots to model news: Sonnet 4.5 launched earlier as a top coding model, and Haiku 4.5 followed shortly after. Haiku 4.5 is framed as delivering similar coding capability with significantly better speed and cost.

    • Sonnet 4.5 positioned as “world’s best coding model” (per video)
    • Haiku 4.5 released soon after as a faster, cheaper option
    • Claimed: similar coding performance to earlier state-of-the-art at ~1/3 cost and >2x speed
    • Noted to surpass Sonnet 4 on some tasks (e.g., computer use)
  6. Haiku 4.5 performance and benchmarks snapshot

    The presenter highlights benchmark numbers to quantify Haiku 4.5’s coding ability and agentic performance. The framing emphasizes that Haiku retains most of Sonnet 4.5’s strength while being cheaper and faster.

    • 73.3% on SWE-bench Verified (as stated)
    • 41% on TerminalBench (as stated)
    • ~90% of Sonnet 4.5 performance in agentic coding evaluations (as stated)
    • Benchmark discussion reinforces the speed/cost/performance tradeoff
  7. When to use Haiku 4.5: plan with Sonnet, implement with Haiku (multi-agent pattern)

    Guidance is given on practical model selection: use Sonnet 4.5 for high-level decomposition and planning, then Haiku 4.5 to implement the code. This approach is described as increasingly popular for large refactors and technical debt work.

    • Sonnet 4.5: higher-level planning and problem decomposition
    • Haiku 4.5: fast implementation and code writing
    • Multi-agent workflows growing in popularity
    • Especially useful for large refactors and technical-debt cleanup
  8. Haiku 4.5 availability and pricing

    Haiku 4.5 is presented as broadly available wherever Claude is accessed, with explicit token pricing. The message reinforces its value proposition for high-throughput coding tasks.

    • Available “wherever you get your Claude” (per video)
    • Pricing: $1 per million input tokens
    • Pricing: $5 per million output tokens
    • Positioned for cost-efficient, fast coding throughput
  9. Returning to the web task: review changes and summaries across backend + frontend

    The demo returns to the running job to inspect outcomes. Claude provides a summary of modifications, including changes in a backend Go service file and frontend TypeScript.

    • Claude produces a concise summary of changes made
    • Backend change example: service.go
    • Frontend change example: TypeScript updates
    • Demonstrates cross-file, cross-layer edits in a single task
  10. “Open in CLI”: teleport from web session to local dev with context preserved

    A highlighted feature lets you continue the exact web session locally via an “Open in CLI” button. You paste a generated command into your terminal to bring the work into your local environment with history, context, and diffs intact.

    • Open in CLI provides a command to run locally
    • Preserves session context: messages, file changes, progress
    • Enables seamless continuation in terminal and VS Code
    • Supports switching between mobile/browser and local workflows
  11. Structured follow-up questions for vague prompts (caching example)

    For broad requests like “add caching,” Claude Code responds with clarifying questions to gather requirements before implementation. The UI walks through choices such as caching level, goals, target area, and technology (e.g., Redis vs in-memory).

    • Claude asks follow-up questions when prompts are underspecified
    • Examples: caching level (DB queries vs API responses)
    • Examples: goals (performance vs DB load) and target area (workouts/users/analytics)
    • Examples: technology selection (Redis, in-memory, Postgres) and submit to implement
  12. Wrap-up: web-based Claude Code, Haiku 4.5 positioning, and availability for Pro/Max

    The video closes by summarizing the main updates: Claude Code on the web for parallel cloud development, Haiku 4.5 for fast/low-cost coding, and improved workflows via structured questions and local handoff. Access is emphasized for Pro and Max users.

    • Claude Code on web: parallel, cloud-based development on the go
    • Haiku 4.5: “lightning-fast” coding at a fraction of the cost (per video)
    • Smarter workflows: multi-question requirement gathering + local handoff
    • Claude Code on the web available now for Pro and Max users at claude.ai/code

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.