AI Benchmarks & Leaderboard — 2026-09-11

New Model Releases & Updates

2 sources

digitalapplied.com

digitalapplied.com

digitalapplied.com

digitalapplied.com

Gemini 3.5 Flash-Lite by Google

• Type: Closed-source, lightweight variant
• Key benchmarks: Improved by 11 Intelligence Index points over predecessor; halves time per task.
• vs. Previous best: Outperforms Gemini 3.5 Flash in intelligence despite being a "Lite" model, while maintaining significantly lower latency.
• What's notable: Demonstrates Google's strategy to maximize intelligence-per-dollar in the small-model category, challenging other "Flash" tier competitors.

Celeris-1 by Unknown Provider (via Artificial Analysis)

• Type: High-speed inference model
• Key benchmarks: 1412 tokens/second throughput.
• vs. Previous best: Currently ranks as the fastest model on the Artificial Analysis leaderboard, surpassing Mercury 2 (744 t/s).
• What's notable: Highlights the growing importance of inference speed as a differentiator for real-time applications like voice agents and live coding assistants.

Mercury 2 by Unknown Provider (via Artificial Analysis)

• Type: High-speed inference model
• Key benchmarks: 744 tokens/second throughput; Intelligence Index competitive with mid-tier models.
• vs. Previous best: Previously held the speed record before Celeris-1's entry.
• What's notable: Remains a top choice for developers balancing speed and intelligence where Celeris-1's specific optimizations may not apply.

Leaderboard Snapshot

Frontier Models (Closed-Source)

Note: Scores vary by leaderboard filter (all models vs. small/medium). Claude Fable 5.1 leads the overall "All Models" leaderboard.

Open-Source Leaders

Note: Open-source rankings are dynamic; DeepSeek V4 and Kimi K3 are frequently cited as the strongest open-weight alternatives to frontier closed models.

Benchmark Deep Dive: The "Lite" Model Paradox

The most striking development this week is the performance of Google's Gemini 3.5 Flash-Lite. According to recent changelog data from Artificial Analysis, the model improved by 11 Intelligence Index points over its predecessor while simultaneously halving the time per task. This is counter-intuitive in a landscape where intelligence typically correlates with parameter count and latency.

The results reveal that architectural efficiency and training data quality are becoming more significant than sheer scale for mid-tier models. Gemini 3.6 Flash, notably, did not improve in intelligence over 3.5 Flash, suggesting that Google's engineering focus has shifted entirely to the "Lite" variants to capture the cost-sensitive developer market. For practitioners, this means the "default" choice for many enterprise applications may shift from standard Pro models to these optimized Lite variants, offering 90% of the capability at a fraction of the cost and latency.

This trend is further supported by the emergence of ultra-fast models like Celeris-1, which hits 1412 tokens per second. The benchmark data suggests that "speed" is no longer just a secondary metric but a primary axis of competition, forcing closed-source providers to optimize for throughput as aggressively as they do for reasoning accuracy.

Analysis & Trends

• State of the art: Anthropic's Claude Fable 5.1 currently holds the #1 spot on the Artificial Analysis LLM Leaderboard with an Intelligence Index score of 53, followed closely by GPT-6 Astra.
• Open vs. Closed gap: The gap is narrowing in specialized domains. Open-weight models like DeepSeek V4 and Kimi K3 are increasingly rivaling closed models on specific benchmarks like SWE-Bench (coding) and MMLU-Pro, though closed models retain an edge in complex, multi-step reasoning tasks.
• Cost-performance: Qwen3.5 0.8B is now the most affordable option at $0.01 per 1M tokens (blended ratio), making it a viable candidate for high-volume, low-complexity tasks previously handled by rule-based systems.
• Emerging patterns: There is a clear bifurcation in model releases: "Max" tier models pushing the ceiling of reasoning (Claude Opus 4.8, GPT-5.5) and "Lite" tier models optimizing for speed/cost (Gemini Flash-Lite, Qwen 0.8B). Mid-tier models are becoming less distinct.

What to Watch Next

• Gemini 3.6 Flash Performance: Monitor if Google addresses the stagnation in intelligence for the non-Lite Flash model in upcoming updates, potentially via a new checkpoint release.
• DeepSeek V4 Official Benchmarks: Look for independent verification of DeepSeek V4's SWE-Bench scores against GPT-5.5 and Claude Opus 4.8 to confirm if open-source has truly caught up in agentic coding.
• Celeris-1 Availability: Track whether Celeris-1's extreme speed (1412 t/s) is available via public API or remains restricted, as this could shift the market for real-time AI agents.