Anthropic Launches Claude Opus 4.8

Anthropic has released its top-tier model, Claude Opus 4.8. Compared to the previous version, 4.7, it improves overall performance across coding, reasoning, and agentic skills, and in particular strengthens Honesty, allowing the AI to flag areas of uncertainty on its own and reducing the probability of missing code flaws by roughly 4x compared to the previous model.

Key Improvements and New Features

• Reliable Agentic Capabilities: Judgment has become more sophisticated during complex multi-step tasks, with enhanced collaboration abilities such as catching its own mistakes or challenging the validity of a plan.
• Cost and Speed Efficiency: 'Fast Mode' is 2.5x faster, while costs are 3x cheaper than the previous model.
• Dynamic Workflows: A research preview feature in Claude Code that can run hundreds of parallel sub-agents to handle complex tasks such as large-scale codebase migrations.
• Effort Control: Users can choose the depth of Claude's thinking, allowing adjustment for either high-quality responses or fast responses depending on the purpose.
• Messages API Update: System entries can now be included within the messages array, allowing instructions to be updated mid-task while preserving the prompt cache.

Future Plans and Pricing Through Project Glasswing, Anthropic is preparing Claude Mythos, a new class of model with intelligence beyond Opus.

Current pricing for Opus 4.8 is as follows.

• Standard Mode: Input $5 / 1M tokens, Output $25 / 1M tokens
• Fast Mode: Input $10 / 1M tokens, Output $50 / 1M tokens