The Claude Opus 4.8 has officially launched, building on the advancements of version 4.7. This new model promises stronger performance across various benchmarks, making it a more effective collaborator for users.
Key Features:
- Control over task effort, allowing users to adjust Claude's performance level.
- Introduction of dynamic workflows in Claude Code for handling large-scale problems.
- Fast mode now operates at 2.5× speed and is significantly cheaper than previous models.
A comparison table highlights the improvements in Opus 4.8 against its predecessor and other models, particularly in coding, reasoning, and practical knowledge tasks. Comprehensive evaluations are detailed in the Claude Opus 4.8 System Card.
User Feedback:
Early testers have noted that Opus 4.8 exhibits greater reliability and improved judgment in agentic tasks. Notably, the model has shown enhanced honesty, being less prone to making unsupported claims. Evaluations indicate that it is four times less likely to overlook flaws in code.
Alignment Assessment:
A thorough alignment assessment prior to release revealed that Opus 4.8 excels in prosocial traits, supporting user autonomy and acting in their best interest. The rates of misaligned behavior are significantly lower than in Opus 4.7, aligning closely with the best-aligned model, Claude Mythos Preview.
Updates and Improvements:
Opus 4.8 defaults to a high effort level, balancing quality and user experience. Users can opt for higher effort levels for more challenging tasks, with increased rate limits in Claude Code to support this. Overall, Opus 4.8 is a modest yet tangible upgrade over its predecessor.
Future Developments:
Plans are underway to release models that deliver similar capabilities at lower costs, alongside a new class of models with even higher intelligence. Currently, select organizations are testing Claude Mythos Preview for cybersecurity applications, with stronger safeguards being developed for broader release.
Pricing:
Claude Opus 4.8 is now available at unchanged pricing: $5 per million input tokens and $25 per million output tokens. Fast mode is priced at $10 per million input tokens and $50 per million output tokens.
Additionally, a new office has been opened in Milan, marking the sixth location in Europe.
