AI Benchmarks & Leaderboard — 2026-10-28

New Model Releases & Updates

GPT-6.1 Terra by OpenAI

• Type: Closed-source mid-tier model
• Key benchmarks: Improved coding and professional task accuracy (specific scores not detailed in source)
• vs. Previous best: Successor to GPT-6.1 Sol; offers 50% lower pricing than the previous model
• What's notable: Enhanced computer-use capabilities targeting routine workplace tasks; part of a broader trend of price cuts by major labs

1 sources

firstpost.com

firstpost.com

Claude Opus 5.5 by Anthropic

• Type: Closed-source frontier model
• Key benchmarks: Artificial Analysis Intelligence Index score of 58 (highest recorded)
• vs. Previous best: Current leader in intelligence, followed closely by Sonnet 5.5 (score 56)
• What's notable: Available in 'max' and 'xhigh' variants; leads in general intelligence metrics

DeepSeek V4.1 Flash 1017 by DeepSeek

• Type: Open-weight model
• Key benchmarks: Notable cost-performance ratio; competitive in open-source leaderboards
• vs. Previous best: Part of China's labs taking the open-weight crown alongside Qwen and Xiaomi's MiMo
• What's notable: Released in October 2026; contributes to the trend of specialized decision AI models

2 sources

local-ai-zone.github.io

local-ai-zone.github.io

local-ai-zone.github.io

local-ai-zone.github.io

Leaderboard Snapshot

Frontier Models (Closed-Source)

Open-Source Leaders

Benchmark Deep Dive

The latest Artificial Analysis leaderboard reveals a tight race for the top intelligence scores. Claude Opus 5.5 (Max) holds the lead with a score of 58, followed by three models tied at 56: Claude Sonnet 5.5 (Max), Claude Opus 5.5 (Xhigh), and Meta's Muse Wonder Full (Max). This clustering at the top suggests diminishing returns in pure intelligence gains for frontier closed-source models, with differentiation now shifting to cost, latency, and specialized capabilities.

In terms of speed, Celeris-1 and K2 Horizon 375B A23B are the fastest models, followed by Mercury 2.5 and Gemini 3.5 Flash-Lite. For lowest latency, North Mini Code and Command A+ lead, followed by NVIDIA Nemotron Nano 9B V2. This indicates a growing specialization where "fast" and "low latency" are becoming distinct metrics for different use cases.

Cost-per-task analysis shows K2 Horizon 375B A23B and Gemma 4 31B as the most efficient, followed by Command A+ and Ling 3.0 Tiny. Meanwhile, Llama 4 Scout and GPT-6 Sol (Daybreak Blue, max) support the largest context windows, with Muse Wonder (max) and Ling 3.1 Flash 1019 close behind. This highlights that context window size is no longer the exclusive domain of the highest-intelligence models.

Analysis & Trends

• State of the art: Anthropic's Claude Opus 5.5 leads in general intelligence, while specialized models like Celeris-1 lead in speed and K2 Horizon in cost-efficiency. OpenAI's GPT-6.1 Terra targets mid-tier reliability and computer-use tasks.
• Open vs. Closed gap: Open-source models like DeepSeek V4.1, Qwen 3.8, and GLM-5.3 are increasingly competitive, particularly in cost-performance ratios, though they still trail Claude Opus 5.5 in pure intelligence scores.
• Cost-performance: OpenAI and Anthropic have cut LLM prices by up to 80% within a few months, making high-performance models more accessible. GPT-6.1 Terra's 50% price cut is part of this aggressive pricing war.
• Emerging patterns: Specialization is key; models are being optimized for specific metrics (speed, latency, cost, context length) rather than just raw intelligence. Chinese labs are taking a strong lead in the open-weight space.

What to Watch Next

• Pricing wars: Expect further price cuts from major labs as OpenAI and Anthropic compete on cost for mid-tier models like GPT-6.1 Terra.
• Open-weight momentum: Watch for new releases from Chinese labs (DeepSeek, Qwen, GLM) that could challenge closed-source models in specific benchmarks.
• Specialized benchmarks: Increased focus on non-intelligence metrics like latency, cost-per-task, and computer-use capabilities as models become more specialized.