Opus 4.5 changed everything (Interview)

Summary of Opus 4.5 changed everything (Interview)

by Changelog Media

1h 44m•February 27, 2026
Listen to Episode

Overview of Opus 4.5 changed everything (Interview)

This episode of The Changelog (host Adam Stachowiak) features Burke Holland (developer relations at GitHub, working across Copilot teams) and centers on how Anthropic’s Opus 4.5 (and the subsequent model race) created a step-change in what developers can build with AI. Burke recounts hands‑on experiments (one‑shoting native apps, replacing SaaS for his wife’s small business), explains current agentic workflows, and discusses the practical and cultural implications for developers, teams, and the software ecosystem.

Episode & guest context

  • Host: Adam Stachowiak (editor-in-chief, Changelog)
  • Guest: Burke Holland (GitHub Copilot team, Developer Relations)
  • Trigger: Burke’s Jan 5 post "Opus 4.5 is going to change everything" and subsequent attention (Hacker News etc.)
  • Themes: Opus 4.5/4.6, GPT-5.3 (Codex), Gemini, Claude Code; Copilot CLI & agent workflows; productivity, craft vs. scale; cost/energy and sustainability; personal software renaissance; "agent psychosis".

Key takeaways

  • Opus 4.5 was a major inflection point: models went from prototyping helpers to actually producing high‑quality, structured code that can one‑shot useful apps.
  • Agentic workflows (planning + autopilot loops + subagents) are becoming the dominant pattern for building nontrivial software with LLMs.
  • Developers still matter: architecture, security, deployment, testing, and production-readiness remain hard and require human oversight and concept-level understanding.
  • The bar for what “good” development looks like has risen—teams that use these tools are much more productive, and that affects expectations in hiring and product shipping.
  • Cost and sustainability are real constraints: current subsidized pricing (discounted tokens, big allowances) won’t last forever; energy/GPU costs will shape future economics.
  • There’s a big opportunity for “personal software” — fast, bespoke tools for individuals or small businesses are easy to build now and flourishing.
  • Psychological risk: “agent psychosis” / addiction — once you open the box, these workflows can consume attention; discipline and boundaries are needed.

Topics discussed (expanded)

  • What changed with Opus 4.5
    • Burke’s experiments: one-shot native Windows tools, a screen-capture→GIF tool, and an iOS app that auto-generates social captions for his wife’s small franchise business.
    • He rebuilt or replaced SaaS solutions locally in hours using modern LLMs.
  • Model landscape
    • Opus 4.5 → Opus 4.6 and 4.6-fast; GPT-5.3 (Codex) described as “a very senior dev”; Gemini also in the mix.
    • Different models have different personalities and strengths — orchestration across models is useful.
  • Workflows & orchestration
    • Plan mode: iterate on requirements with the agent to surface missing constraints.
    • Autopilot / Ralph loops: agents run iterative passes until they reach confidence.
    • Anvil (Burke’s custom agent): classifies task difficulty (easy/medium/hard) and routes work accordingly; uses subagents, multiple models, and verification/skills.
    • Verification is critical: agents must produce verifiable evidence their work works (tests, browser checks, etc.).
  • Economics & sustainability
    • Holiday/limited-time increased quotas drove huge experimental usage (Claude Code example).
    • Subsidized token pricing enables rapid experimentation but is unsustainable unless costs drop or business models adjust.
  • Developer roles & learning
    • Coding is less the bottleneck; shipping and production are the hard parts.
    • AI accelerates conceptual learning (ETL patterns, networking, protocols) and flattens language/tool boundaries for builders.
    • The suggestion that “everyone will be a builder” — developers will shift, but craftsmanship and high-end engineering remain valuable.
  • Open source & incumbents
    • Open source is thriving; cheap creation lifts more projects and tooling.
    • Incumbents have advantages (ubiquity, models trained on them), but novel, compatible alternatives can still emerge.
  • Craft vs. scale
    • Concern about losing craftsmanship (analogy: cathedrals vs. cheap mass-produced homes). There will be room for both mass-produced and “artisanal” software, likely at different price points.
  • Human factors
    • “Agent psychosis”: the mental/attention cost of running and supervising agents — it can be addictive and intrusive.
    • Discipline, prioritization, and boundaries become more important for builders.

Notable quotes & paraphrases

  • “Opus 4.5 changed everything.” — framing the episode
  • “If AI isn't useful, it's your problem, not the AI.” — on personal responsibility to experiment and shape workflows
  • “The world really is a developer's oyster.” — on the opportunity right now
  • “We need people who understand architecture, security, and how apps get to production.” — why developers still matter
  • “Agent psychosis” — the phenomenon of being consumed by agent-driven workflows

Practical recommendations / action items

  • Experiment now:
    • Try Copilot CLI (GA announced in episode) and agentic features in your editor/CLI.
    • Test different models (Opus 4.6, GPT-5.3, Gemini, Claude) to understand strengths and how to orchestrate them.
  • Start small, plan first:
    • Use "plan mode" or an iterative planning step to surface missing requirements before hands-off runs.
  • Build personal software:
    • Identify routine tasks or paid SaaS you use and test replacing them with bespoke tools (for private use) to learn.
  • Implement verification:
    • Ensure agents produce evidence (tests, automated browser checks, logs) so you can trust outputs.
  • Watch costs and limits:
    • Begin with conservative plans, monitor token/usage, and build fail-safes (rate limits, budget alerts).
  • Protect your attention:
    • Set boundaries for agent notifications, daily standups/check-ins with agents, and prioritize finishing projects, not just starting them.
  • Learn concepts over syntax:
    • Focus on learning system design, architecture, and conceptual patterns (ETL, security, deployment) — these scale with AI assistance.

Sponsors & links mentioned in episode

  • Changelog community (Zulip): changelog.com/community
  • Changelog Plus Plus: changelog.com/plusplus
  • Augment Code / Augie (CLI): augmentcode.com
  • Squarespace (promo): squarespace.com/changelog
  • Notion (custom agents): notion.com/changelog
  • Card My Yard (example small business): cardmyyard.com
  • Mentioned projects & resources: Copilot CLI (GitHub), OpenNext, diffs.com / pierre.computer (example new diff renderer), Tigris (tigrisdata.com)

Closing thought / perspective

The episode’s core thesis: modern LLMs (Opus 4.5/4.6, GPT-5.3, Gemini, Claude) have shifted the practical baseline for what a single developer or small team can build quickly. That creates huge opportunity (personal software, faster learning, more builders) but also raises real questions about production quality, costs, sustainability, and preserving craftsmanship. The pragmatic advice is: experiment, build workflows that include planning and verification, watch your costs, and be mindful of human attention and process as agents become part of everyday software development.