Claude Opus 4.8
Maximum reasoning for complex, long-horizon product work.
Previous generation. Use Claude Opus 5 instead. Moved to Anthropic's legacy list on 30 Jul 2026. Still runs, but Opus 5 costs the same and is more capable.
- Maker
- Anthropic
- Context
- 1M tokens
- Pricing
- $5 / 1M tokens / $25 / 1M tokens
Identical pricing to Opus 5, which supersedes it. No reason to pick this.
Best for
- Deep reasoning over messy, multi-source product context
- Long-horizon agentic work (multi-step research, doc generation)
- High-stakes synthesis where nuance matters more than speed
Prompting quirks
Rewards explicit structure with XML-style tags
Wrap distinct inputs in tags like <context>, <transcript>, <task>. Claude follows the boundaries you draw far more reliably than prose separators — and it cuts re-reads, so you spend fewer tokens.
Tell it the role and the stakes, not just the task
"You are reviewing this PRD for an exec audience; flag anything that wouldn't survive a CFO question" outperforms "review this PRD."
Ask for the thinking, then the answer
Let it reason in a <scratchpad> before the final output when the task is genuinely hard. For simple tasks, skip it — it just burns tokens.
Copy-paste example
Tuned for Claude Opus 4.8. Swap in your own content where marked.
<role>You are a seasoned CPO reviewing a PRD before it goes to the board.</role>
<prd>
{paste your PRD here}
</prd>
<task>
Identify the three weakest assumptions. For each: why it's risky, the
one question an investor would ask, and how I'd de-risk it in a week.
Be blunt. Bullet points only.
</task>Facts verified against the official Anthropic docs on 19 August 2026.