Claude Opus 5: The Daily Driver Gets a Brain Transplant

Fable-level reasoning on software engineering tasks for about half the token cost—and the model checks its own work instead of handing you broken code.

Anthropic dropped Opus 5 today. Same price as Opus 4.8 ($5/$25 per million input/output tokens), but Frontier-Bench scores more than doubled. On CursorBench 3.2 at max effort, it lands within half a percent of Fable 5's peak while burning half the tokens per task. That's your CI budget breathing easier.

What Actually Changes Tomorrow

Not the benchmarks—how the model behaves when nobody's looking.

Testers keep using the same word: judgment. One engineer gave Opus 5 a market data feed for a new exchange. Previous models couldn't finish even with detailed plans. Opus 5 built its own test harness because there was no live feed to validate against. That's being careful.

Another test: debugging a real bug in a popular open-source package manager. Opus 5 found an edge case the community patch missed. A competing model fixed only the surface symptom. Same pattern: thoroughness over speed.

The Practical Trade-offs

Same safeguards as Opus 4.8, with slightly tighter guardrails on vulnerability scanning and exploit generation. If your workflow involves penetration testing, you'll hit the classifiers—though Anthropic claims 85% fewer interventions than Fable 5, and flagged requests fall back to Opus 4.8 by default.

Cybersecurity remains Mythos 5's territory. Opus 5 is competitive at finding vulnerabilities but substantially behind at turning them into working exploits. Stick with Mythos for that.

For everyone else: the consistency improvement alone might justify the switch. Lovable reported 22% gains on agentic coding tasks over Opus 4.7, but variance dropped sharply. Reliable output, build after build—the kind of win that compounds.

The Boring Stuff That Matters

Fast mode runs at ~2.5x default speed for twice the base price. Available on the Claude Platform and through usage credits in Claude Code. Two beta features worth noting: mid-conversation tool changes without cache invalidation, and automatic API fallbacks that route around safety classifier blocks instead of hard-failing.

Opus 5 is the new default on Claude Max and the strongest option on Claude Pro. API endpoint: claude-opus-5.

If you're shipping code tomorrow, this is the model you want running your PRs and debugging sessions—not because it's the smartest, but because it double-checks before saying 'done.'