Sign in Subscribe
news

2026-10-04 Claude Opus 5.5 - Give it the whole task, keep your hand on the brake

Sebastien Dubois

Sebastien Dubois

— 5 min read

Canonical version: 2026-10-04 Claude Opus 5.5 - Give it the whole task, keep your hand on the brake.

Anthropic published a post on getting the most out of Claude Opus 5.5 in Claude and Claude Code. It reached 229 points on Hacker News on 3 October. I found the comments more useful than the post: people tried the advice on real work and reported what broke.

In this piece, I go through both, then give my own recommendations.

What the tips get right

The core idea is the one from my Claude Opus 5.5 note: hand the model a whole task and let it run.

It uses adaptive thinking, meaning it "decides how carefully it needs to think after each question". The old "think carefully" incantations mostly add noise now.

The post also asks you to define the task clearly and say what "done" means. That's one sentence, and it's what keeps an agent from going on without knowing when to stop (the most expensive failure in agentic work, in my opinion).

On parallel subagents, the post says they work with little oversight. One commenter has numbers to back it: Opus 5.5 plus subagents cut a CI pipeline from about 10 minutes to about 4, and billing minutes dropped by about 60%. It took 12 PRs over a 9-hour session, "less than an hour of my attention", and roughly $500 worth of tokens on the $100 plan.

Cache reads now cost 5% of the input price, down from 10%. Agents reread the same big context again and again, so that price cut hits them directly.

Other commenters were positive too. One says it beats Claude Fable 5.1 on coding tasks in VS Code. Several mention faster token generation than OpenAI's Sol 6.1, better spatial reasoning for frontend work and blueprints, and a big step up from the "writing slop" of Opus 5. Another had it read vintage circuit schematics from service manuals; it cross-checked datasheets, verified the physics and caught typos in the originals. There are also a few "three months of work in one week" claims, which I'd take with a grain of salt.

Where the pushback is valid

"Don't ask for reasoning" is too blunt

The post advises against asking for visible reasoning in outputs. One commenter calls that misleading: asking it to "think through this step by step" still surfaces how parts of a task depend on each other, and "different issues emerge" when you do. In their view, planning is still a skill you have to learn.

I agree. Dumping the model's internal reasoning into the reply is what the reasoning_extraction safeguard flags. A written plan you review before it starts is something else, and I'd keep asking for one. Commenters also split the work between models: Fable 5.1 for high-level planning and code review, Opus 5.5 for "sterile execution". That split makes sense to me.

Autonomy cuts both ways

Long autonomous runs get risky on infrastructure. One commenter saw it expand a deployment to 5 regions without warning, and it argued when corrected. Another asked it to summarize a README; it executed untrusted code and solved math problems nobody asked for.

The README story worries me more. A model that acts on what it reads in a file, instead of describing it, is open to Prompt injection, whatever the alignment numbers at launch say.

There's also the opposite failure. Linas Beliūnas points out that it sometimes ends a turn with a progress report. If your agent loop reads that report as "done", it stops halfway through. So the finish line has to be something you can check (tests, a checklist), not the model's last message.

Fast changes, slow maintenance

Someone pointed out that AI-optimized CI ends up "full of cludges", with huge inline bash scripts in workflow YAML files. Another said a mid-level engineer would spot those inefficiencies anyway, and the maintenance burden rarely justifies the gain for a team. A third ran architecture work over 3 iterations and more than a million tokens, and got overcomplicated designs and security errors that needed heavy human review.

A 60% cut in CI minutes saves money, but only if someone on the team can maintain what the agent wrote. Otherwise it turns into Technical debt.

Independent checks point the same way. Sonar found 44% more concurrency problems per line in Opus 5.5's code than in Opus 5's. Endor Labs found that only 33.5% of its security fixes passed their checks once memorized answers were excluded. It's smarter (it ranks first on the Artificial Analysis Intelligence Index), but that doesn't make its code safe to merge unread.

Fan-out costs real money

The subagent advice says little about token costs. On High effort, one user's schematics work ate about half of his weekly budget. Another commenter hesitates to move to the $200 plan because he gets billed for thinking tokens that produce no output. And the $500 worth of tokens on a $100 plan only works because the subscription absorbs it; on the API, you'd pay the $500.

Effort makes it worse. At max, Opus 5.5 produces about 1.6× as many output tokens as Opus 5, so the cost per task ends up the same as Opus 5's. The "40% cheaper" only holds at lower effort, and medium already matches or beats Opus 5 at high.

Cyber safeguards still misfire

People doing legitimate security work say the safeguards make it nearly useless. One was reverse-engineering an early-2000s disc cloning tool. The classifier got "poisoned" and terminated every single response, workarounds included. Another cites "spurious cyber refusal" as a reason not to upgrade. Add the routing detail from launch week (flagged cybersecurity requests get rerouted to an older model) and on security topics, you don't fully control which model answers you. And according to Linas Beliūnas, since 24 September the API bills you for refusals of requests to show its reasoning.

What I'd do

I'd still follow the post (describe the whole task once, then let it work), with these limits.

Keep the finish line, and add stop lines. "Done means the tests pass", plus "stop and ask before touching DNS, regions, IAM or anything that costs money".

For multi-step work, ask for the plan as a file, review it, then let it execute. If you have Fable access, use it for the plan and the review, and Opus 5.5 for the execution.

Keep permission prompts on for infrastructure and destructive commands, and scope the credentials the agent can reach. A rule in CLAUDE.md is guidance; a permission prompt actually stops it.

Also:

  • Treat every file it reads as untrusted. Summarizing a README doesn't need shell access, so give read-only tasks a read-only setup
  • Start fan-out at medium effort, cap the number of subagents, and check your weekly usage before you switch to High. Avoid max: it eats the whole cost advantage over Opus 5
  • Don't let an agent loop treat the model's last message as "done". Check the finish line yourself
  • Review AI-made CI and infra changes like any PR. If you'd reject a 200-line inline bash script from a colleague, reject it from Claude too
  • If you do security work, expect refusals, and look at the Cyber Verification Program if you qualify

The full breakdown (pricing, benchmarks, API changes, safeguards, prompting tips) is in my Claude Opus 5.5 note.

That's it for today! ✨

References

About Sébastien

Ready to get to the next level?

Found this valuable? Share it with someone who needs it.

Join 6,000+ readers. Get practical systems for knowledge & AI. Free.

Subscribe ✨

Free: Knowledge System Checklist

A clear roadmap to building your own knowledge system. Subscribe and get it straight to your inbox.

6,000+ readers. No spam. Unsubscribe anytime.

Subscribe