The current era of large language model development has shifted from a race for raw parameter count to a battle over cognitive depth. For months, the developer community has watched as models evolved from simple next-token predictors into reasoning engines capable of internal deliberation. This week, the industry reached a new benchmark of clarity as the latest Intelligence Index v4.1 results revealed exactly where the ceiling of machine intelligence currently sits and which models are actually pushing it higher.
The Architecture of Intelligence and the New Leader
Claude Opus 5, specifically in its Adaptive Reasoning Max Effort configuration, has claimed the top spot among 170 evaluated models with a score of 61. This result places it at the absolute peak of the current intelligence hierarchy. The evaluation was particularly rigorous because it focused heavily on a subgroup of 126 reasoning models. Unlike standard LLMs that generate responses instantaneously, these reasoning models employ extended thinking processes, allowing the system to iterate on a problem internally before committing to a final output. The data shows that this capacity for expanded thought is the primary driver behind the high scores in the Intelligence Index.
To understand the weight of this 61-point score, one must look at the composition of the Intelligence Index v4.1. This metric is not a single test but a composite of nine distinct, high-difficulty evaluations designed to stress-test different cognitive dimensions. These include GDPval-AA v2, τ³-Banking, Terminal-Bench v2.1, SciCode, Humanity's Last Exam, GPQA Diamond, CritPt, AA-Omniscience, and AA-LCR. By aggregating performance across agentic tasks, complex coding, scientific reasoning, knowledge reliability, and long-context inference, the index converts multifaceted capabilities into a single, comparable intelligence value.
While Claude Opus 5 dominates the intelligence metric, other models have pushed the boundaries of data ingestion. Llama 4 Scout emerged as the leader in memory capacity, offering a massive 10 million token maximum context window. In practical terms, the context window defines the total sum of input and output tokens a model can handle in a single session. While the output limit is typically much smaller than the total window, a 10M token capacity fundamentally changes the scale of Retrieval-Augmented Generation (RAG) workflows, allowing the model to reference vast libraries of documentation without losing the thread of the conversation.
The Trade-off Between Sovereignty, Speed, and Cognition
When the data is analyzed beyond the top rank, a clear tension emerges between three competing priorities: the intelligence of the model, the speed of the response, and the level of control over the weights. The most striking revelation is the persistent gap between closed-source proprietary models and open-weight alternatives. Out of the 170 models tested, 94 were open-weight models, meaning their internal parameters are available for public deployment. Among these, GLM-5.2 (max) performed the best, scoring 51 points. While impressive, this creates a 10-point intelligence deficit when compared to the top-tier closed models. For enterprises, this represents a critical strategic crossroads: whether to accept a lower cognitive ceiling in exchange for the security and customization of self-hosted infrastructure, or to rely on the superior intelligence of a closed API.
Furthermore, the results highlight that intelligence is not a monolithic trait but a configurable setting. The performance of the Claude family demonstrates a hierarchical structure based on effort levels. Claude Opus 5 (Adaptive Reasoning, Max Effort) hit 61, while the Xhigh Effort version and Claude Fable 5 (Adaptive Reasoning, Max Effort, Opus 4.8 Fallback) both followed closely at 60. Meanwhile, GPT-5.6 Sol (max) and the High Effort version of Claude Opus 5 both landed at 59. This suggests that the same underlying model can be tuned to trade off computation time for accuracy, allowing developers to choose the specific level of reasoning effort required for a given task.
This trade-off becomes even more apparent when comparing intelligence to latency. The models at the top of the Intelligence Index are not the fastest. For instance, Mercury 2 recorded a blistering output speed of 901.6 tokens per second, making it the fastest text generator in the test group. Similarly, Gemini 2.5 Flash-Lite (Non-reasoning) achieved a time-to-first-token (TTFT) of just 0.34 seconds, minimizing the perceived lag for the end user. The data proves that high intelligence and real-time responsiveness are currently inverse variables. A model that spends more time in the extended thinking phase to achieve a score of 61 cannot simultaneously offer the sub-second latency of a Flash-Lite model.
Ultimately, the Intelligence Index v4.1 reveals that the market has fragmented into specialized tiers. There is the cognitive peak occupied by Max Effort reasoning models, the high-throughput tier led by Mercury 2, the massive-context tier defined by Llama 4 Scout, and the sovereign tier led by GLM-5.2. The decision for a technical lead is no longer about finding the best model, but about determining which of these three constraints—intelligence, speed, or control—is the one they can afford to sacrifice.
Strategic AI deployment now requires a precise calculation of the performance-cost threshold where additional intelligence no longer justifies the increase in latency or the loss of infrastructure control.
