AI Benchmarks & Leaderboard — 2026-04-24

New Model Releases & Updates

GPT-5.5 by OpenAI

• Type: Closed-source
• Key benchmarks: Intelligence Index score of 60 (xhigh setting), 59 (high setting) — highest on Artificial Analysis leaderboard
• vs. Previous best: Surpasses GPT-5.4 (score: 57) and Claude Opus 4.7 (score: 57), previously tied for first
• What's notable: The xhigh compute setting pushes it clearly ahead of all rivals on Artificial Analysis's composite Intelligence Index; GPT-5.3 Codex (xhigh) also appears in the top 5 with a score of 54, indicating OpenAI is occupying multiple top spots simultaneously

Kimi K2.6 by Moonshot AI

• Type: Closed-source
• Key benchmarks: Intelligence Index score of 54, placing it 4th on the Artificial Analysis leaderboard
• vs. Previous best: Enters the top 5, matching GPT-5.3 Codex (xhigh)
• What's notable: Represents continued strong performance from Chinese AI labs in frontier model rankings; sits just below the OpenAI/Anthropic/Google top tier

GLM-5.1 by Zhipu AI

• Type: Open-source
• Key benchmarks: SWE-Bench Pro score of 58.4%, leading all open-source models on that coding benchmark
• vs. Previous best: Tops the open-source coding leaderboard per April 2026 tracking
• What's notable: Zhipu's GLM family continues to lead open-source coding benchmarks; also compared favorably against Gemma 4, Qwen 3.6, Llama 4, and DeepSeek V4 in April rankings

Leaderboard Snapshot

Frontier Models (Closed-Source)

Open-Source Leaders

Benchmark Deep Dive

Open-Source AI's Strategic Moment: Closing the Gap on Closed Models

A Forbes analysis published April 19, 2026 argues that the question is no longer whether open-source AI matters, but whether companies can still afford to treat it as secondary. This framing resonates with the hard benchmark data emerging this week.

The April 2026 open-source rankings tell a striking story. GLM-5.1 now scores 58.4% on SWE-Bench Pro — a demanding software engineering evaluation designed to resist saturation — putting it ahead of Gemma 4, Qwen 3.6, Llama 4, and DeepSeek V4. For context, SWE-Bench Pro tasks models with resolving real GitHub issues, making it a meaningful proxy for production coding utility rather than synthetic test performance.

Speed metrics add another dimension. Mercury 2 clocks 716.5 tokens per second on the Artificial Analysis leaderboard, followed by Granite 4.0 H Small at 453.0 t/s and Granite 3.3 8B at 379.5 t/s. These are open-weight models matching or exceeding the inference speed of many closed APIs. Meanwhile, Qwen 3.5 0.8B holds the cost crown at $0.02 per million tokens — a price point that makes experimentation effectively free for most organizations.

For practitioners, the practical implication is clear: the open-source tier has matured to the point where, for coding, speed-sensitive, or cost-constrained workloads, closed frontier models are no longer the default answer. The intelligence gap at the very top (GPT-5.5 at 60 vs. GLM-5.1's open-source leadership) remains real, but the functional gap for everyday deployment has narrowed substantially.

Analysis & Trends

• State of the art: GPT-5.5 leads the closed-source composite Intelligence Index (score: 60). For coding, GLM-5.1 leads open-source SWE-Bench Pro at 58.4%. For speed, Mercury 2 (716.5 t/s) dominates inference throughput across all model categories.

• Open vs. Closed gap: The gap at the absolute frontier persists — GPT-5.5 at 60 is meaningfully ahead of the best open-source models in composite reasoning. However, in specific verticals like coding (GLM-5.1) and cost-efficiency (Qwen 3.5 0.8B at $0.02/M tokens), open-source options are now enterprise-grade alternatives. The Forbes analysis from April 19 formalizes what practitioners have been observing for months.

• Cost-performance: Qwen 3.5 0.8B holds the most affordable position at $0.02 per million tokens (blended), followed by Gemma 3n E4B Instruct at $0.03. These prices represent a continued commoditization of capable small models.

• Emerging patterns: Speed differentiation is becoming a new competitive axis — Mercury 2 at 716.5 t/s is nearly 60% faster than the third-place Granite 3.3 8B (379.5 t/s). Chinese AI labs (Moonshot's Kimi K2.6 at #4 closed-source, Zhipu's GLM-5.1 leading open-source coding) continue to punch well above their weight in frontier rankings.

What to Watch Next

• GPT-5.5 vs. Claude Opus 4.7 head-to-head on domain-specific benchmarks: The composite Intelligence Index has GPT-5.5 ahead, but Claude Opus 4.7 with Adaptive Reasoning may close or reverse that gap on reasoning-heavy or long-context tasks. Watch for third-party evaluations breaking down performance by task type.

• GLM-5.1 SWE-Bench Pro trajectory: If GLM-5.1 continues improving on coding benchmarks, it could become the de facto open-source default for software engineering agents — a category with significant commercial implications.

• Enterprise open-source adoption signals: With Forbes now explicitly framing open-source AI as a strategic imperative rather than a niche choice, watch for enterprise procurement and deployment data that tests whether this thesis holds in practice — particularly as IBM's Granite 4.0 family shows strong speed numbers that could appeal to on-premises deployments.