Opus 4.6 Outshines Competitors with Higher Intelligence Yield and Lower Compute Requirements

In the world of AI, where everyone is vying for the most advanced models, a key metric that often gets overlooked is the Intelligence Yield (IY). This metric not only measures how hard a task a model can solve but also how efficiently it does so. The latest data from various frontier models shows that Anthropic's Opus 4.6 stands out by solving harder tasks more reliably with significantly less compute.

Understanding Intelligence Yield

Intelligence Yield (IY) is a crucial metric for evaluating AI models, particularly in terms of efficiency and effectiveness. It quantifies how much "work" a model can do relative to the amount of compute it requires. An IY of 1.0 means that one minute of human work is equivalent to one minute of compute time. Higher values indicate more efficient models.

Intelligence Yield by Model

Let's dive into the latest figures:

• Opus 4.6 (Anthropic): IY = 3.00

  • Solves tasks up to 14.5 hours in complexity.
  • Achieves a 78.9% average score.
  • Requires only 3 hours and 49 minutes of compute time.

• Opus 4.5 (Anthropic): IY = 0.71

  • Solves tasks up to 5.3 hours in complexity.
  • Achieves a 73.0% average score.
  • Requires 5 hours and 29 minutes of compute time.

• Opus 4 (Anthropic): IY = 0.25

  • Solves tasks up to 1.7 hours in complexity.
  • Achieves a 61.6% average score.
  • Requires 4 hours and 6 minutes of compute time.

• Codex 5.3 (OpenAI): IY = 0.22

  • Solves tasks up to 6.5 hours in complexity.
  • Achieves a 74.5% average score.
  • Requires 21 hours and 59 minutes of compute time.

• Gemini 3 Pro (Google): IY = 0.19

  • Solves tasks up to 3.9 hours in complexity.
  • Achieves a 71.0% average score.
  • Requires 14 hours and 29 minutes of compute time.

• GPT-5.2 (OpenAI): IY = 0.035

  • Solves tasks up to 6.6 hours in complexity.
  • Achieves a 75.3% average score.
  • Requires 142 hours and 24 minutes of compute time.

• GPT-5 (OpenAI): IY = 0.029

  • Solves tasks up to 3.6 hours in complexity.
  • Achieves a 69.4% average score.
  • Requires 86 hours and 43 minutes of compute time.

• o3 (OpenAI): IY = 0.014

  • Solves tasks up to 2.0 hours in complexity.
  • Achieves a 63.6% average score.
  • Requires 90 hours and 41 minutes of compute time.

Intelligence Yield Over Time

The trend over time is also revealing. The chart below shows the release date versus IY for each model, highlighting the rapid advancements in AI efficiency:

• 2025: Models like Opus 4, GPT-5, and Gemini 3 Pro set the stage with modest IY values.
• 2026: Opus 4.5 and Codex 5.3 show significant improvements, but it's Opus 4.6 that truly stands out.

Key Takeaways

1. Efficiency Gains: Opus 4.6 demonstrates a remarkable improvement in efficiency, achieving an IY of 3.00. This means it can solve tasks that would take over 14 hours for a human to complete using only about 3.5