The Opus 4.8
Cheat Sheet

Source: Anthropic: Claude Opus 4.8 →

Anthropic shipped Claude Opus 4.8 on May 28, 2026. It's the newest, most capable Claude. Three things genuinely changed: effort control finally landed in the regular Claude app and in Cowork (it was already in Claude Code), the model got measurably more honest, and Claude Code can now spin up hundreds of subagents in a single session via dynamic workflows.

This guide is the cheat sheet for the first two. The dynamic-workflows feature gets its own walkthrough (linked at the bottom). Below: what effort actually does, when to pick each level, how to use the honesty upgrade so you stop double-checking everything, and the prompts that land harder on 4.8 than they did on 4.7.

Effort is a setting that tells Claude how hard to think. Turn it up, Claude takes longer and gives a more careful answer. Turn it down, you get a faster answer for simple work. The default jumped from medium-ish to high on 4.8, which means if you don't touch it, every prompt now silently spends more compute than before.

Here's how to actually decide.

· Low — one-off rewrites, quick formatting, short summaries, anything you'd skim in two seconds. Faster, cheaper, totally fine.

· Medium — emails you'd want a second read on, content drafts, normal coding tasks. Good baseline if you'd rather spend less and only crank it up when you notice the output isn't good enough.

· High — analysis, anything multi-step, code that has to actually work, anything where a better answer saves you real time. This is now the default on 4.8.

· xhigh (Claude Code only, via ultracode) — the deep-thinking tier. Big migrations, hard debugging, the kind of work where you'd happily wait a minute for a better answer. Burns way more tokens, so reach for it on purpose.

The pattern that works: start at high. Drop to medium when you notice you're paying for thinking you don't need. Bump to xhigh when the task is genuinely hard. Never default to xhigh on chat-style work — it's a money trap.

Anthropic's headline claim: 4.8 is more honest. Independent reviewers confirmed it — on every benchmark, 4.8 had the lowest incorrect-answer rate, mostly because it abstains when uncertain instead of confidently making something up. For code review specifically, it's about 4x less likely to let a flaw in its own code slip past without flagging it.

This is the upgrade you actually feel in real work. But you only feel it if you ask the right way.

Three small prompt tweaks that unlock it:

· Ask for the uncertainty. Instead of "give me the answer," say "give me the answer and flag anything you're not sure about with a confidence score." 4.8 actually does it.

· Ask it to refuse. Add "if you don't know, say you don't know. Do not guess." 4.8 will hold the line where 4.7 would have improvised.

· Ask it to mark hallucination-risk content. Especially useful for research, legal, medical, anything cite-able. "Mark any specific fact (number, date, quote, name) that you'd want me to verify before using."

If you're calling Claude via the API or Anthropic SDK, three things changed in 4.8 that will break old code or save you money if you adapt:

· Temperature / top_p / top_k are gone. 4.8 doesn't accept them. If your code passes them, you'll 400-error on migration. Steer via prompting instead.

· Mid-conversation system messages no longer break prompt cache. This is the big money win for anyone running long agentic loops. You can inject new instructions mid-chat and the cache survives. Simon Willison called this the real unlock.

· Refusal categories are now public via stop_details. When 4.8 refuses, your app can read why and route around it instead of retrying blindly.

Opus 4.8 is the most capable but also the most expensive Claude. The mistake most people make is using it for everything — it's a money leak and the speed cost adds up too. Here's the calling order I use:

· Haiku 4.5 — classification, extraction, formatting, anything you'd batch through. Roughly 3x faster and dramatically cheaper than Sonnet.

· Sonnet 4.6 — your daily driver. Trails Opus by about 1 point on SWE-bench but is way cheaper. Use it for 80–90% of tasks.

· Opus 4.8 — novel problem-solving, multi-step reasoning, anything where honest abstention matters (research, legal, medical), or where a better answer saves real money.

· Opus 4.8 fast mode — only when wall-clock speed matters more than cost. About 2.5x faster output, costs more per token but Anthropic noted it's 3x cheaper than fast mode used to be on prior models.

Anthropic is positioning 4.8 as a sharpening of 4.7, not a generational leap. The benchmarks are real (4x less likely to miss code flaws, 84% on Online-Mind2Web, top scores on agentic coding), but you probably won't feel a night-and-day difference on quick chat-style work.

Some honest pushback from the community in the first week:

· Sycophancy regression on r/ClaudeAI. Some users are reporting 4.8 opens with "you're right to…" more often than 4.7 did. Counter the prompt with: "do not flatter me or agree by default. Push back if I'm wrong."

· Hallucinated file paths in Claude Code launch-day threads. A handful of users had sessions burn tokens on No such file or directory loops. The fix: name files with full paths in your prompt and let Claude verify the path exists before editing.

· Prompt-injection guardrails are slightly weaker on browser-use tasks. Zvi Mowshowitz flagged a regression on VendBench. If you're running browser agents unsupervised on the public web, don't auto-execute trades or purchases from 4.8 output without a human gate.

My take so far: the honesty win is real and shows up in research and code work. The effort dial is the actually-useful UX change. Most of the rest is incremental. Use it, but don't expect magic.