AI Model Leaderboard: Intelligence, Speed, and Price Comparison

Artificial Analysis evaluates hundreds of AI models using a proprietary Intelligence Index and various performance metrics like speed and latency. Claude Opus 5 and GPT-5.6 Sol lead in intelligence, while models like Mercury 2 and Gemini Flash-Lite offer superior speed and low latency. The platform helps developers choose models by visualizing the trade-offs between cost, quality, and technical capabilities.

Key Points

• Claude Opus 5 and GPT-5.6 Sol are the top-performing models in terms of intelligence and complex reasoning.
• Mercury 2 and Gemini 3.5 Flash-Lite are the industry leaders for output speed and low-latency performance.
• The Intelligence Index v4.1 uses a multi-faceted evaluation approach including agentic work, coding, and scientific reasoning.
• Open-weights models are competitive, with GLM-5.2 (max) currently ranked as the most intelligent non-proprietary model.
• Pricing and cost-per-task vary significantly based on token types, including input, output, and prompt caching discounts.

Sentiment

The community is mixed and analytical, with a skeptical tilt toward leaderboard claims rather than toward the underlying data itself. Many commenters agree that comparative benchmarks are useful inputs, especially when they include price and speed, but they reject treating the top intelligence score as a definitive model-selection answer. The strongest consensus is that users should match models to specific tasks and workflows rather than chase a universal winner.

In Agreement

• The leaderboard is useful because it brings together intelligence, speed, cost, latency, and other comparative data that practitioners can use as inputs when selecting a model.
• Aggregate benchmarks are imperfect but still better than relying only on anecdote; broad rankings can correlate with practical capability when interpreted carefully.
• The highest-ranked models do matter for difficult tasks where smaller or cheaper models can stall, loop, or fail despite being adequate for easier work.
• Cost and speed comparisons are especially valuable because the best practical model is often the one that reaches sufficient quality at lower latency or lower spend.
• The presence of multiple dimensions and specialization benchmarks makes the site more useful than a simple one-number ranking.

Opposed

• A single intelligence leaderboard encourages misleading claims that one model is universally best, even though models have jagged strengths across domains and tasks.
• Leaderboards can become more useful for vendor bragging rights than for end-user decisions unless they reflect specific workloads and real production behavior.
• Many common tasks are already saturated, making smaller, faster, cheaper, or open-weight models better practical choices than the top-ranked frontier model.
• Early leaderboard wins for newly released models are noisy and may reflect hype, benchmark targeting, or first-impression bias rather than durable usefulness.
• Benchmarks often fail to capture important workflow factors such as overthinking, context degradation, tool-use behavior, coding style, and how well models collaborate or critique each other.