Claude Opus 4.8 Cheat Sheet: Effort Levels, Honesty Mode, and the Prompts That Land

Effort tells Claude how hard to think. Higher means slower, more careful answers and more of your usage. Lower means faster answers for simple work. In the Claude app, click the model name next to the send button, then Effort, and pick a level. You can change it mid-chat and it applies from the next reply. (Claude Help Center)

Low

Quick rewrites, formatting, short summaries, anything you’d skim in two seconds.

Medium

Emails, content drafts and everyday tasks. A good baseline if you’d rather save usage and only go up when the answer isn’t good enough.

High

Analysis, multi-step work, code that has to work, anything where a better answer saves you real time. High was the default on Opus 4.8.

Extra high and Max

Extra high is built for long coding and agent tasks and is available on Opus 4.7 and newer. Max is for problems where correctness matters more than speed. Both burn through usage fast, so pick them on purpose.

What Changed On 5.5

On Opus 5.5, thinking is always on and effort is your only control. Its API default dropped to medium. In the app, the menu marks each model’s default, so check it before a hard task and bump it up yourself.

Anthropic says Opus 4.8 is "around four times less likely than its predecessor to allow flaws in code it has written to pass unremarked," and testers said it flagged uncertainty more often. Anthropic calls Opus 5.5 its "strongest model on most measures of honesty." You get the most out of that when you ask for it directly.

Ask for the uncertainty

Instead of "give me the answer," say "give me the answer and flag anything you’re not sure about."

Give it permission to say I don’t know

Add "if you don’t know, say so. Do not guess." It’s the fastest fix for confident wrong answers.

Make it mark facts to check

For research, legal, medical or anything you’ll cite, ask it to tag every number, date, quote and name you should verify. I paste this at the top of every research chat.

Honesty Ground Rules

You are working with me on [research / a draft / an analysis] about [topic]. Before we start, here are the ground rules for this whole conversation. 1. If you don't know something, say "I don't know." Do not guess or fill gaps with something plausible. 2. Tag every specific fact you state (number, date, quote, name, statistic, citation) with one of: [SOLID] you are confident and it is widely documented [LIKELY] probably right, but I should spot-check it [VERIFY] could be wrong, check it before I use it 3. When you are interpreting or giving an opinion instead of stating a fact, start that sentence with "My read:". 4. If a question needs current information, say so and either search the web or tell me exactly which source would answer it. 5. Do not flatter me or agree by default. If my premise is wrong, say so first. 6. End every answer with a one-line confidence summary: what's solid, what's shaky, and what to double-check first. Confirm these rules in one sentence, then answer my first question: [your question]

Make it review its own work

Use this before you ship code, send a report or trust an analysis.

Pre-Ship Review

You are a senior reviewer checking work before it ships. Review the [code / document / analysis] I attach or paste below. What it is supposed to do: [one or two sentences] Who it's for: [audience or users] What "done" means: [tests pass / numbers tie out / ready to send to a client] Steps: 1. Restate in two sentences what this is trying to do, so I can confirm you understood it. 2. List every problem you find, ranked by severity (blocker, should fix, nice to have). For each: where it is, what's wrong, and the fix. 3. Flag anything you are unsure about separately, and say what would settle it (a test to run, a source to check, a question for me). 4. If you wrote or changed anything earlier in this chat, review your own work with the same standard and call out your own mistakes. Rules: don't rewrite the whole thing. Don't pad the list with style nitpicks if there are real problems. If you find nothing serious, say so plainly instead of inventing issues. [Paste or attach the work here]

Sampling settings are gone. Opus 4.8 rejects temperature, top_p and top_k with a 400 error, and the newest API version rejects them on every model. Steer with your prompt instead.
Mid-task instructions keep your cache. Since 4.8, you can add a system entry inside the messages list to update instructions without breaking the prompt cache. Big savings on long agent loops.
Refusals tell you why. When Claude declines, stop_reason is refusal and stop_details gives a category, so your app can route around it instead of retrying blindly.

Moving to Opus 5.5 (claude-opus-5-5)? Thinking can’t be turned off, so remove any disabled-thinking setting and use effort instead. Set effort explicitly, since the default is now medium. Forced tool choice returns an error, so use auto plus a clear instruction. (Opus 5.5 migration guide)

Using the top model for everything wastes money and time. Here’s the current lineup with API prices per million input and output tokens. (Anthropic models overview)

Haiku 5.5 (from $0.10 / $0.50) for sorting, extracting, formatting and anything you run in bulk.
Sonnet 5.5 ($2 / $10) as your daily driver. Fast and smart enough for most tasks.
Opus 5.5 ($4 / $20) for long agent work, hard reasoning and anything where an honest "I’m not sure" matters. Anthropic suggests starting here if you’re unsure.
Fable 5.1 ($10 / $50) for the hardest reasoning and long-running projects, when Opus 5.5 at higher effort still falls short.
Fast mode on Opus 5.5 runs up to 2.5 times faster at double the price. Use it only when speed matters more than cost.

For the full rundown, see my guides on which model to use and Opus vs Fable.

The Opus 4.8Cheat Sheet

The Opus 4.8
Cheat Sheet