Claude Opus 5 is available today across all platforms, delivering meaningful benchmark improvements over Opus 4.8 without any price increase. Anthropic is keeping the same $5 per million input tokens and $25 per million output tokens, which makes this a straight-up efficiency win for anyone already paying for Opus-tier access.
What Opus 5 Actually Improves
Anthropic describes Opus 5 as a model built for daily use that approaches the intelligence of Claude Fable 5 at half the price. Positioning Opus 5 as near-peer to a frontier model is a significant claim. Opus 5 outperforms Opus 4.8 across several categories:
- Agentic coding
- Novel problem solving
- Knowledge work
- Computer use
- Multidisciplinary reasoning
- Scientific research and visual output quality
The model is also better at verifying its own work and iterating until a task is complete, which matters for agentic workflows where human oversight is limited. On cybersecurity, Opus 5 is close to Mythos 5 at finding vulnerabilities, though it remains substantially behind Mythos 5 on exploitation, partly because it was not specifically trained on cyber tasks.
A Fast mode is available at 2.5x the default speed for twice the base price, for situations where latency matters more than cost.
Opus 5 is now the default model on Claude Max and the strongest available model on Claude Pro. That default status means the average Claude Max subscriber gets a better model today without changing any settings.
Two developer-facing updates shipped alongside Opus 5. First, mid-conversation tool changes – developers can now swap which tools Claude can access during an active conversation without invalidating the prompt cache. This is a practical improvement for anyone building complex agentic pipelines where tool availability changes based on context. Second, automatic fallbacks – API requests flagged by safety classifiers on Opus 5 or Fable 5 can now automatically reroute to another model instead of being blocked outright. With fallbacks enabled, requests always go to the best available model rather than failing silently.
