AI Benchmarks & Leaderboard — 2026-09-03

New Model Releases & Updates

Muse Spark 1.3 by Meta

• Type: Closed-source (implied by "released" context, though Meta often releases open weights; specific parameter count not disclosed in snippet)
• Key benchmarks: Chief AI Officer Alexandr Wang claimed it surpasses OpenAI's GPT-5.6 Sol in coding capabilities.
• vs. Previous best: Directly challenges GPT-5.6 Sol, currently a top-tier competitor in coding tasks.
• What's notable: This release represents Meta's most powerful AI model to date, intensifying competition in the coding and agent task sectors.

Leaderboard Snapshot

Frontier Models (Closed-Source)

Open-Source Leaders

Benchmark Deep Dive

The latest updates from Artificial Analysis highlight a tightening race between frontier closed-source models and high-performing open-weights models. Claude Fable 5.1 currently leads among 150+ reasoning models with an Intelligence Index score of 66. This score aggregates performance across 10 challenging evaluations, providing a robust measure of general intelligence rather than just specific task proficiency.

A notable trend is the cost-performance ratio. While Claude Fable 5.1 leads in raw intelligence, Granite 4.2 3B has emerged as the most cost-efficient model, with the lowest cost per Intelligence Index task at $0.0047. This suggests that for practitioners where budget is a primary constraint and absolute state-of-the-art reasoning is not required, smaller, optimized models are becoming increasingly viable.

The gap between the top open-weights model, Kimi K3 (max), and the top closed-source model, Claude Fable 5.1, is now just 6 points on the Intelligence Index (60 vs. 66). This narrow margin indicates that open-source models are reaching parity with closed-source counterparts for many enterprise use cases, particularly when combined with specialized fine-tuning or retrieval-augmented generation techniques.

Analysis & Trends

• State of the art: Claude Fable 5.1 leads in general reasoning and adaptive effort scenarios. For coding, Meta's newly released Muse Spark 1.3 claims to outperform GPT-5.6 Sol, suggesting a potential shift in leadership for developer-focused tasks.
• Open vs. Closed gap: The gap is narrowing significantly. Kimi K3 and GLM-5.3 are within striking distance of the top closed-source models on the Intelligence Index, offering competitive performance with greater flexibility for deployment.
• Cost-performance: Granite 4.2 3B and GPT-5.6 Luna (low) are leading in cost efficiency, making them attractive for high-volume, lower-complexity tasks. The cost per Intelligence Index task is a key metric for practitioners balancing budget against capability.
• Emerging patterns: There is a clear trend towards "adaptive reasoning" models like Claude Fable 5.1 and Opus 4.8, which allow users to balance latency and cost by adjusting reasoning effort levels.

What to Watch Next

• Independent Verification of Muse Spark 1.3: Third-party benchmarks will be crucial to validate Meta's claim that Muse Spark 1.3 surpasses GPT-5.6 Sol in coding tasks.
• Open-Weights Momentum: Continued improvements from Kimi K3 and GLM-5.3 could see them overtake current closed-source leaders in specific niche benchmarks or cost-adjusted intelligence metrics.
• Gemini 3.7 Flash Performance: Recent changelogs indicate Gemini 3.7 Flash improved 4 points over its predecessor, potentially reshuffling the mid-tier leaderboard if it becomes widely available.