Source: raw/Using_Claude_Opus_5.5_as_your_daily_driver.md — “Using Claude Opus 5.5 as your daily driver”, Anthropic’s official Claude YouTube channel (@claude), https://www.youtube.com/watch?v=jKRl_CSVxyI, uploaded 2026-09-23, 3:33. The presenter is not named. Upload date, duration and description from ai-research/anthropic-youtube-model-launch-video-descriptions-2026-10-30.md.

Anthropic’s own walkthrough of Opus 5.5 in everyday Claude Code work, recorded the day after launch. It runs Opus 5 and Opus 5.5 side by side on the same bug, shows what each does to a Max plan’s five-hour limit, and gives three working rules: stay on medium effort, go to high only for changes that span many files, and run read-only subagents on a smaller model. All measurements are the presenter’s own runs, and the presenter says yours will differ.

Key Takeaways

  • Anthropic’s pitch, in the video’s words. Compared with Opus 5, Opus 5.5 is “faster, it’s a lot better at telling you what it did, and it’s 20% cheaper per token.” On Pro, Max or Team “your limits go 25% further,” and “because it needs fewer tokens to get to a result, a task is 40% cheaper.” The 20% and 40% figures match the launch page. The launch page says five-hour limits went up but gives no percentage, so the 25% figure appears only in this video.
  • Side-by-side bug fix, same prompt, both on medium effort. The bug: refunds on orders with a discount code came out a few cents off. Both models read the refunds code, found where the discount is split, fixed it and ran the tests. Opus 5.5 finished in under a minute, and its whole reply fit on one screen: what it changed, the tests it ran, “the one thing it ran into, and the one question it has for me.” Opus 5 first read the README, the money helpers and the test suite, made more tool calls and wrote a longer reply. In that run Opus 5.5 was 30% faster.
  • Limit use on a Max plan: about 4% vs 6%. Opus 5.5 used about 4% of the five-hour limit and Opus 5 about 6%. The presenter ran it “a bunch of times and every time 5.5 used less of my 5-hour limit than Opus 5 did.”
  • Medium effort is the default, and enough for most daily work. The video defines effort as “how much thinking it puts into a turn before it acts.” Everything in the bug-fix demo ran on medium.
  • Go to high for changes that span many files, then come back. The demo: on medium, asked to rename a field in the orders handler, Opus 5.5 renamed it there and the tests passed. But the mobile app sends the same field as customer_ref through a serializer, so a search for customerRef never found it. “That fix isn’t wrong, it’s just not entirely complete.” On high, told to finish the rename everywhere, it traced what sends that payload, found the mobile client and the serializer, renamed the field there too, and added a test so the miss would not recur. The rule: high for a fix that spans “multiple files or the entire project”, then back to medium.
  • Run read-only subagents on Sonnet, sometimes Haiku. “A subagent that’s only exploring the codebase doesn’t need Opus-level reasoning.” Set the model in a custom subagent’s frontmatter, or set CLAUDE_CODE_SUBAGENT_MODEL in settings.json as the default for all of them.
  • Run prompt-audit on your CLAUDE.md and skills. The video names “/claude-api prompt-audit”, which “goes through your CLAUDE.md and your skills and tunes the instructions in there for Opus 5.5”, and says it is useful even if you only use Claude Code.

The daily-driver recipe

The video’s closing summary, as a checklist:

SituationSetting
Most daily workOpus 5.5, medium effort
A change across many files or the whole projectHigh effort for that change, then back to medium
Subagents that only read or exploreSonnet, “and sometimes even Haiku”
After switching to Opus 5.5Run prompt-audit over CLAUDE.md and skills

How it fits the rest of the wiki

  • The effort rule matches Anthropic’s other guidance. Thariq Shihipar’s tests found effort mattered most on tasks with many edge cases and least on builds from a detailed spec; see Which Effort Level to Use. The rename demo is an edge-case miss of the same kind.
  • CLAUDE_CODE_SUBAGENT_MODEL is a default, not an override, since v2.1.251 (2026-08-28). A model: line in an agent definition and an explicit per-spawn model both take precedence over it. CLAUDE_CODE_SUBAGENT_MODEL_FORCE (v2.1.257) applies it to every subagent regardless. Details in Subagents.
  • Two prompt-audit commands exist. The claude-api skill has a prompt-audit subcommand that dates each instruction, and Claude Code v2.1.283 (2026-09-25) added a built-in /doctor prompt-audit that audits CLAUDE.md files, skills, agents and commands for patterns written for older models. See Agent Skills. The video names the skill version, two days before the built-in one shipped.

Try It

  1. Check your effort setting. In Claude Code, run /effort and confirm Opus 5.5 is on medium. Since v2.1.280, an effort level saved before /effort became per-model no longer applies to new models such as Opus 5.5 (see Model vs. Effort).
  2. Repeat the side-by-side on one of your own bugs. Run the same prompt on Opus 5 (/model claude-opus-5) and Opus 5.5 from the same starting state, and note time, tool calls and the change in your five-hour limit.
  3. Raise effort only for cross-cutting changes. Before a rename, schema change or API change, switch to high. When the change is merged, switch back to medium.
  4. Move explorer subagents off Opus. Add model: sonnet to the frontmatter of each custom subagent that only reads code, or set CLAUDE_CODE_SUBAGENT_MODEL once in settings.json.
  5. Audit your instructions. Run /doctor prompt-audit, or the claude-api skill’s prompt-audit subcommand, and fix what it flags first.

Implementation

Tool/Service: Claude Code with Claude Opus 5.5 (claude-opus-5-5)

Setup: Default Opus in Claude Code from v2.1.280 (2026-09-22). Default effort is medium.

Cost: 20 output / $0.20 cache read per MTok on the API. On subscriptions, cost shows up as share of the five-hour and weekly limits.

Integration notes: The subagent default goes in the env block of settings.json. The video names the variable but does not show the file, so the exact value below is an assumption.

{
  "env": {
    "CLAUDE_CODE_SUBAGENT_MODEL": "sonnet"
  }
}

Per-agent override, in a custom subagent’s frontmatter (.claude/agents/<name>.md):

---
name: code-explorer
description: Read-only codebase exploration
model: sonnet
---

Open Questions

  • What “limits go 25% further” measures. The video does not say whether the 25% is the raised five-hour cap, the effect of cheaper tokens, or both. The launch page gives no percentage for the limit increase.
  • One presenter’s runs. The 30% speed gap and the 4% vs 6% limit use come from the presenter’s own Max-plan runs. The video says the result held “every time” but gives no run count or task list.
  • Which prompt-audit. The video says “/claude-api prompt-audit”. That reads as the claude-api skill’s subcommand, but a transcript cannot show what was typed on screen, so it is not confirmed.