How Anthropic Engineers Actually Use Claude Opus 5.5
What this video is
⚡ a 7-minute video, readable in 60 secondsAnthropic released Claude Opus 5.5, announced Sep 22, 2026, with a playbook guide titled "Getting the most out of Opus 5.5 in Claude and Claude Code." The headline change is that Opus 5.5 generates output over 30% faster than Opus 5, with a price drop and increased five-hour usage limits on Pro, Max, Team, and seat-based Enterprise plans, plus new default behaviors (working longer autonomously, thinking before every reply). No breaking changes or required migration were stated; Sonnet 5.5 and Haiku 5.5 are coming in the following weeks but had not shipped as of this release. On-screen release notes and code were not captured, so this synopsis is built from narration only.
Release: Claude Opus 5.5, described as the most powerful model on the planet, released [00:08]
Key takeaways
- Release: Anthropic playbook "Getting the most out of Opus 5.5 in Claude and Claude Code," published Sep 22, 2026, 9 minute read [00:00]
- Changed: Opus 5.5 generates output more than 30% faster than Opus 5 [00:14]
- Changed: Price drop and increased five-hour usage limits on Pro, Max, Team, and seat-based Enterprise plans [00:14]
- Roadmap (not yet shipped): Claude Sonnet 5.5 and Claude Haiku 5.5 coming in the following weeks with many of the same performance, efficiency, and safety improvements [00:14]
+ 24 more takeaways
- Benchmark: Opus 5.5 scores 66.4% on Agentic coding versus Fable 5.1's 55.8% [00:15]
- Benchmark: Opus 5.5 outperforms GPT-6 Astra and GPT-5.6 Sol at a fraction of the cost, per Anthropic's benchmarks [00:15]
- Added: Opus 5.5 leads in agentic coding, computer use, and knowledge work benchmarks, per Anthropic [00:15]
- Changed: Opus 5.5 works longer on its own, plainly tells you what it did, and thinks before every reply, unlike prior usage patterns [00:38]
- Changed: Opus 5.5 keeps going on long, multi-part work better than Opus 5, with biggest gains on multi-step tasks like carrying a change through a large repo until tests pass [02:05]
- Added: Early testers ran Opus 5.5 for hours with little oversight [02:05]
- Tip 1: Hand over the whole task and state what "done" looks like, then let Opus 5.5 work [00:51]
- Tip 1 detail: Give the whole task in one message and name the finish line, such as "the tests pass" or "every endpoint is migrated," then let it run [02:21]
- Tip 2: Delete "think carefully" lines since Opus 5.5 already thinks before every reply [00:51]
- Changed: You no longer need to tell Opus to "think hard," "think carefully," or "think step by step" because it does this automatically [03:04]
- Changed: Anthropic testing found removing "think carefully" lines made replies start sooner with no clear drop in quality [03:11]
- Endpoint/Control: In Claude Code, change the effort slider to adjust how much Opus 5.5 thinks; say "answer directly" for a quick answer to a simple question [03:25]
- Tip 3: When a long run ends, read what Claude needs from you first before responding [01:55]
- Added: You can type a follow-up mid-run to inject input into a running Opus 5.5 task without waiting for the current response to finish [03:41]
- Changed: Waiting for Claude to finish before following up makes it re-ingest all context, increasing tokens and cost [03:48]
- Changed: On Opus 5.5, runs are longer, so a restart costs more, making mid-run follow-ups more valuable [03:39]
- Tip: For design work, name the specific styles you don't want rather than giving a vague instruction, since a general instruction like "avoid a generic look" mostly swaps one default style for another on Opus 5.5 [04:15]
- Tip: For steering a long Claude Code run, put a short rule in CLAUDE.md about when to stop and ask versus when to keep going [04:57]
- Migration/Workflow: Suggested CLAUDE.md rule separates KEEP GOING (steps that don't need input, status notes in the same message as the next action) from STOP AND ASK (only when work can't continue without the user or before destructive actions like deleting data, force-pushing, or changing anything outside the repo) [05:25]
- Changed: Replying "continue" works when a run stops with "Want me to continue?", but adding the CLAUDE.md rule reduces how often that happens [05:59]
- Tip: Explicitly ask Claude to split big work across subagents, especially for large audits or migrations with little oversight [06:02]
- Benefits: Splitting work across subagents makes tasks run faster and saves tokens [06:17]
- Tip: Keep the task list in a file such as TASKS.md rather than relying on Claude's own memory or context window [06:23]
- Benefits: A task list kept in a file survives context compaction on long runs and shows at a glance what's done and what's left [06:48]
Shown on screen — grab and go
PROMPTThe payment-endpoint migration prompt
Migrate the payment endpoints from the old client to the new one. Done means: every endpoint uses the new client, the old client is deleted, and the test suite passes. Stop and ask me only if a test fails for a reason you can't explain.
shown fully legible in a code block on screen at 02:49 shown at 2:49
PROMPTThe 'no generic look' website design prompt
Build a personal website with placeholder content. Don't use a cream or off-white background, italic accent words in headings, numbered "01 / 02 / 03" section labels, monospace labels, or pill-shaped buttons.
shown fully legible in a code block on screen at 04:35 shown at 4:35
CODECLAUDE.md rule for when to stop and ask
## KEEP GOING When a step doesn't need my input, keep going. Put status notes in the same message as your next action. ## STOP AND ASK Stop and ask only when you can't continue without me, or before anything destructive: deleting data, force-pushing, or changing anything outside this repository.
reconstructed from the on-screen CLAUDE.md code block at 05:36, merged with the surrounding legible read at 05:25 shown at 5:36
PROMPTSubagent audit instruction for retry bug
Audit every service in services/ for the retry bug in the linked issue. Give each service to its own subagent. When a subagent reports back, check its evidence before you accept it. Finish with one table: service, affected yes or no, and the evidence.
transcribed from the screen at 06:11-06:13, fully legible shown at 6:11
PROMPTTask-list-in-a-file instruction
Keep a checklist in TASKS.md. Tick each item when it's done, and add anything new you find.
transcribed from the screen at 07:12, fully legible shown at 7:12
How this brief was shaped: Changelog / Release Notes · confidence Medium
Transcript and OCR both center on the Opus 5.5 release itself, price drop, usage limit increases, benchmark table, and communication improvements, framed as a playbook of best practices for using the new model in Claude Code. The OCR shows an official announcement-style benchmark table and feature list rather than live code, while the transcript gives numbered practical tips for working with the new model.
The lens sets this brief's structure, never its facts — every claim is held to the same citation and fact-check standard.
Jump to a moment
Their links, sorted & clickable
🏛️ Communities & courses1
👉🔗 Other links1
👉Was this brief useful?