<aside> <img src="/icons/reorder_gray.svg" alt="/icons/reorder_gray.svg" width="40px" />

Table of Contents

</aside>

【0.】What changed in Opus 4.8

Anthropic shipped Claude Opus 4.8 on May 28, 2026. They call it a “modest but tangible” upgrade — and on raw benchmarks they’re right (GPT-5.5 still wins on terminal work). The real shift is quieter: the model now tells you when it isn’t sure. It’s about 4× less likely than 4.7 to let a flaw in its own code pass without flagging it.

【1.】The 4 Upgrades That Matter

<aside> 🔹

1. Honesty — it stops bluffing

• What it means: Around 4× less likely than 4.7 to let a flaw in its own code pass unremarked. It flags what it’s unsure about instead of guessing.
• Why it matters: If you ship work with AI, this is the upgrade that counts.
• The hack: Ask Claude to rate its own confidence and list its assumptions — see Prompt 1. </aside>

<aside> 🔹

2. Effort control — you set how hard it thinks

• What it means: A control next to the model picker: low / medium / high / xhigh, plus max. Default is high.
• Why it matters: You trade tokens for quality on purpose now.
• The hack: High covers ~90% of work. Use xhigh for hard problems; max uses ≈4× the thinking tokens — save it for code you’re about to ship. See Prompts 3–4. </aside>

<aside> 🔹

3. Dynamic workflows — many subagents, one run

• What it means: In Claude Code, it can plan a job, run many parallel subagents in one session, and verify its own output — even codebase-scale migrations.
• Why it matters: Less hand-holding on big jobs.
• The hack: Hand it the goal + success criteria and let it plan, split, run, self-check. Max/Team/Enterprise only, and it burns tokens fast — start scoped. See Prompts 5–6. </aside>

<aside> 🔹

4. Fast mode — faster, and cheaper than it used to be

• What it means: ~2.5× faster, and 3× cheaper than the old fast mode (still ~2× the standard price: $10 / $50 per million vs $5 / $25).
• Why it matters: Great for quick drafts and iteration.
• The hack: Don’t use it for production code review or architecture calls — fast-mode accuracy isn’t publicly benchmarked. </aside>

【2.】The Numbers

Source: Anthropic, “Introducing Claude Opus 4.8,” May 28, 2026. Benchmarks are Anthropic’s own.

<aside> ▪️

Where it still falls short (the honest part)

1. Incremental. Anthropic itself calls it “modest but tangible.” Happy on 4.7? No need to rush.
2. GPT-5.5 still leads on terminal-heavy coding, advanced math, and web research.
3. Prompt-injection resistance regressed slightly at the model level vs 4.7; product safeguards mostly compensate.
4. Easy to overspend — a single max-effort + dynamic-workflow session can eat days of a small plan’s budget.
5. Fast mode has no published accuracy benchmarks. Don’t use it for mission-critical work. </aside>

【3.】10 Mega Prompts for Opus 4.8