Version 11

Since mid-June 2026, four open-weights models — GLM-5.2 (Zhipu AI), Ornith-1.0 (DeepReinforce), VibeThinker-3B (Weibo AI), and Agents-A1 — have posted benchmark results closing part of the gap with leading closed systems on coding, math, and abstract reasoning. GLM-5.2 leads open-weights models on DeepSWE[1] and holds the highest verified open-weights score on ARC-AGI-2 at 22.8% versus GPT-5.5's 85%[2], but whether its scores reflect genuine capability or distillation from Claude Opus remains disputed[8][9]. ARC Prize has announced ARC-AGI-3 as the next benchmark in the series[17], with an associated competition and arxiv paper now public[18], as ARC-AGI-2 scores rose from 3% in May 2025 to 85% for GPT-5.5. Separately, Fable leads KernelBench-Mega over Claude Opus 4.8 and GPT 5.5[16], and ByteDance Seed's EdgeBench finds top frontier models roughly doubling their in-context learning speed on 12–72 hour tasks every three months[15].

Open-weights models reaching frontier-adjacent scores on domain-specific benchmarks offer developers cheaper, inspectable alternatives to proprietary APIs, but the distillation question is unresolved and affects how reliably those scores generalize. The gap between high static benchmark scores and low agentic performance — Claude Opus 4.8 at only 20.6% on OSWorld 2.0 long-horizon tasks[16] — suggests benchmark rankings are an incomplete proxy for production utility.

Since mid-June 2026, a cluster of open-weights models have produced benchmark results that close some of the gap with leading closed frontier systems on coding, mathematical reasoning, and abstract visual tasks. GLM-5.2 from Zhipu AI — a ~750B total / ~40B active MoE architecture — leads open-weights models on the DeepSWE software engineering leaderboard[1] and has been verified by ARC Prize at 22.8% on ARC-AGI-2 at $0.25 per task, the highest publicly reported open-weights score on that benchmark, versus GPT-5.5's 85%[2]. VibeThinker-3B from Weibo AI scores 94.3 on AIME26 and 80.2 Pass@1 on LiveCodeBench v6 from only 3 billion parameters under MIT license[3]. DeepReinforce's Ornith-1.0-397B (17B active parameters) scores 82.4 on SWE-Bench Verified[4]; Simon Willison tested its 35B GGUF locally and found proficient agentic performance at 103 tokens per second under Apache 2.0[5]. Agents-A1 (35B) trains on 45K-token verified multi-step trajectories and distills specialist teacher models, claiming 1T-parameter-model performance under Apache 2.0[6].

GLM-5.2 has attracted the most concentrated analysis and the sharpest disagreement. Zvi Mowshowitz called it the best available open-weights model and a potential 'DeepSeek moment' for open-source agents[7], but simultaneously argued it is heavily distilled from Claude Opus — citing its tendency to self-identify as Claude and its use of a Claude-based evaluation harness — and commercially awkward for most production use[8]. An independent investigation by Lisan al Gaib into GLM-5.2's post-training methods found the results credibility-questionable, reinforcing the distillation concern[9]. Against this skepticism, hands-on evidence has been consistently positive: bendee983, normally skeptical of benchmarks, found it genuinely impressive in direct use[10]; Semgrep reports it beating Claude in their internal cybersecurity benchmarks[11]; Julian Goldie SEO reports it winning 4 out of 5 real-world builds against Claude Opus 4.8[12]; and developers report cost savings using it in Cursor[13]. Google's July 8 Android Bench update adds GLM-5.2 to that domain-specific coding benchmark alongside seven other models — and notes that Gemini lags behind the competition on Google's own benchmark[14].

The broader benchmark picture extends beyond open-weights comparisons. ByteDance Seed's EdgeBench evaluates agents on 134 real-world tasks requiring 12 to 72 hours of interaction, measuring in-context experiential learning rather than one-shot reasoning[15]; top frontier models appear to be roughly doubling their 2-hour learning speed every three months, a potential scaling dynamic that most static benchmarks do not capture. Jack Clark's Import AI newsletter reports that Fable achieved an 18.71X speedup on KernelBench-Mega, outperforming Claude Opus 4.8 (14.4X) and GPT 5.5 (4.34X)[16]; Clark's Remote Labor Index shows AI success rate on online freelance tasks rose from 2.5% to 16.1% between October 2025 and July 2026, and OSWorld 2.0 data shows Claude Opus 4.8 at only 20.6% binary accuracy on long-horizon computer tasks.

ARC Prize has announced ARC-AGI-3 as the next benchmark in its series, with milestone prizes, an active competition, and a published arxiv paper[17][18]. ARC-AGI-2 scores rose from 3% in May 2025 to 85% for GPT-5.5, compressing its discriminative value; ARC-AGI-3 is positioned to evaluate frontier agentic intelligence at a higher difficulty level. A calibration analysis had used GLM-5.2's verified ARC-AGI-2 score to estimate closed model capabilities from private benchmark data[19][20], but if distillation from closed models partially explains GLM-5.2's scores, those estimates carry systematic bias — a question that will persist regardless of which benchmark series takes precedence.