Claude Opus 5.5 takes top spot in new AI intelligence index

A new ruler for AI power

Artificial Analysis recently released Intelligence Index v5, a major update to its widely-cited benchmark for measuring artificial intelligence. This version represents the most significant overhaul to the index since its original launch. The update changes how the industry measures the capabilities of the world's most advanced models.

The new index is designed to be harder for developers to game. According to Artificial Analysis, 60% of the score now relies on private test sets that AI labs have never seen before. This shift is intended to provide a more accurate picture of how models perform on novel tasks rather than on data they might have encountered during training.

Because of these changes, the leaderboard has shifted significantly. Claude Opus 5.5 now sits at the top of the rankings with a score of 58. It narrowly leads OpenAI's GPT-6 Astra, which scored 57, and GPT-6.1 Sol, which scored 55. These three frontier models are separated by only two points total, showing how competitive the top tier remains.

The open-weight divide

While the top of the leaderboard remains crowded, the new methodology has created a much larger distance between private models and open-weight ones. Under the previous version of the index, the gap between the best closed model and the best open-weight model was only 3 points. Under the new v5 rules, that gap has widened to 12 points.

Kimi K3.1 Pro from Moonshot AI currently holds the lead for open-weight models with a score of 46. This is a notable reversal from a few months ago when open-weight models were closing the distance to the leaders. Previously, the gap had been cut roughly in half over a single year, but the new private test sets appear to be hitting open models harder.

Analysts suggest this might be because open-weight labs have fewer resources to build private evaluation partnerships. Frontier labs with deeper benchmarking relationships may have more access to held-out data. Whether this reflects a true lack of capability or just a difference in how the tests were built is currently a major point of debate in the AI community.

Technical specs of Kimi K3.1 Pro

Moonshot AI developed Kimi K3.1 Pro as a mixture-of-experts model, which is a technique where only specific parts of the model activate for different tasks. This specific model features a total of 2.8 trillion parameters, though only 104 billion are active at any one time. It is designed to handle a 1 million token context window, which refers to the amount of information the model can process at once.

The company also released a faster variant called Kimi K3.1 Pro Highspeed. This version is already accessible to users through Moonshot's existing applications and API. These technical choices allow the model to balance high-level reasoning with the speed required for practical use cases.

The performance of these models can be fluid depending on the benchmark used. For example, Meta's Muse Wonder Full scored 56 on the old index just two weeks before the new version launched. This highlights how quickly the competitive landscape moves and how much a single score can change based on the underlying methodology.

Impact on enterprise buyers

Artificial Analysis does not sell models, which makes its index a valuable tool for corporate procurement teams. Companies often use these independent scores to justify spending on AI infrastructure. When a business is defending a seven-figure contract to a board, a neutral composite score carries more weight than a lab's own marketing materials.

However, the new index creates a difficult situation for buyers who already made decisions based on the old scores. If a company chose a specific vendor because it appeared close to the leaders on the old benchmark, the math has changed overnight. The new index effectively bakes in a dozen-point penalty for open-weight models that wasn't there before.

To help with this, Artificial Analysis launched a companion tool called the Progress Index v1. This tool places every major language model released since November 2022 on a single capability scale. It allows buyers to see not just who is winning today, but how quickly the ground is shifting as new models arrive every few months.