Claude Opus 4.8 is here. Is it as good as they say?

What It Covers

Claire Vo shares early hands-on testing of Anthropic's Claude Opus 4.8, a coding-focused agent model priced at $5/$25 per million tokens, evaluating its performance across greenfield coding, existing codebases, and business strategy tasks.

Key Questions Answered

• •Greenfield vs. existing code: Opus 4.8 performs well on one-shot, net-new feature builds — it planned and autonomously coded a full prototyping tool in roughly 20 minutes — but degrades significantly when navigating existing codebases, rebasing branches, or resolving edge-case bugs.
• •Hallucination risk under confidence: Despite running on high-effort mode, Opus 4.8 fabricated conclusions from hypotheses rather than validated data, both in coding and strategy contexts. Treat high-confidence outputs with skepticism and explicitly prompt it to verify sources before accepting results.
• •Strategy work: 4.7 outperforms 4.8: Side-by-side testing on a business strategy prompt showed Opus 4.7 anchored responses in specific numbers and structured data, while 4.8 produced vague, hand-wavy roadmaps. For data-driven strategy tasks, 4.7 remains the stronger choice.
• •New agentic infrastructure worth testing: Claude Code now supports dynamic workflows enabling hundreds of parallel sub-agents. Claude.ai and CoWork gain effort control settings from low to max. These harness-level changes may offset model limitations when prompting strategies are tuned appropriately.