Version 9

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 produced benchmark results narrowing the gap with leading closed systems on coding, math, and abstract reasoning[1][4][3][6]. GLM-5.2, a ~750B total / ~40B active MoE architecture, leads open-weights models on DeepSWE and holds a verified 22.8% on ARC-AGI-2 at $0.25 per task, far below GPT-5.5's 85% but the highest publicly reported open-weights score[2]; whether those scores reflect genuine capability or training on closed-model outputs remains actively disputed[8][9][11]. Separately, Fable — a newer closed model — achieved an 18.71X speedup on KernelBench-Mega, outperforming Claude Opus 4.8 (14.4X) and GPT 5.5 (4.34X)[17], while ByteDance Seed's EdgeBench documents frontier models doubling their in-context learning speed on 12–72 hour tasks every three months[16].

Open-weights models reaching frontier-adjacent benchmark scores offer developers cheaper, inspectable alternatives to proprietary APIs — but whether those scores reflect genuine generalization or distillation from closed models is unresolved. Fable's KernelBench-Mega lead over both Claude and GPT on GPU kernel optimization is notable because, as Jack Clark argues, autonomous kernel development feeds directly into AI systems' ability to improve their own training infrastructure[17]. OSWorld 2.0 data showing Claude Opus 4.8 at only 20.6% on long-horizon computer tasks[17] and EdgeBench's findings on in-context learning both suggest that static benchmark scores and agentic real-world performance remain substantially decoupled.

Since mid-June 2026, four 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 3 billion parameters under MIT license[3]. Ornith-1.0-397B from DeepReinforce — 397B total, 17B active — 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 drawn the most concentrated analysis and the sharpest disagreement. Zvi Mowshowitz called it the best available open-weights model and framed it as 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, amplifying rather than resolving the distillation concern[9]. Against this skepticism, hands-on evidence is 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 build comparisons against Claude Opus 4.8[12]; and developers report meaningful cost savings using it in Cursor[13]. A calibration analysis used GLM-5.2's verified ARC-AGI-2 score to estimate closed model capabilities from eight private benchmark datapoints[14][15] — an approach whose reliability depends directly on whether GLM-5.2's scores are genuine or artifacts of distillation.

The benchmark picture has additional dimensions beyond open-weights models. 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[16]; top frontier models appear to be roughly doubling their 2-hour learning speed every three months, a potential scaling law that most benchmarks used to rank open-weights models do not capture. Jack Clark's Import AI newsletter reports that Fable achieved an 18.71X speedup on KernelBench-Mega using CUDA on an RTX PRO 6000 Blackwell, outperforming Claude Opus 4.8 (14.4X) and GPT 5.5 (4.34X)[17] — a result Clark frames as significant because autonomous kernel development is a prerequisite for AI systems to improve their own training infrastructure. Clark's Remote Labor Index separately shows AI success rate on online freelance tasks rose from 2.5% to 16.1% between October 2025 and July 2026[17], and OSWorld 2.0 data shows Claude Opus 4.8 at only 20.6% binary accuracy on long-horizon computer tasks with a 1.6-hour median human completion time — a contrast to high single-task benchmark scores that parallels EdgeBench's findings on the gap between static and experiential evaluation. ARC Prize has announced ARC-AGI-3 milestone prizes[18], indicating the abstract reasoning benchmark series is advancing past ARC-AGI-2, whose scores rose from 3% (May 2025) to 85% for GPT-5.5.