Version 8

Since late June 2026, four open-weights models — GLM-5.2 (Zhipu AI), Ornith-1.0 (DeepReinforce), VibeThinker-3B (Weibo AI), and Agents-A1 (35B) — have produced benchmark results narrowing capability gaps with frontier closed systems on coding, math, and abstract reasoning[2][5][4][7]. GLM-5.2, a ~750B total / ~40B active MoE architecture, leads open-weights systems on DeepSWE and holds a verified 22.8% on ARC-AGI-2 at $0.25 per task — well below GPT-5.5's 85%[3][1]. The distillation question — whether GLM-5.2's scores reflect genuine capability or training on closed-model outputs — remains actively contested: an independent post-training investigation found results "sus"[10] while Semgrep reports GLM-5.2 beating Claude in cybersecurity benchmarks[12]. ByteDance Seed's EdgeBench introduces a separate evaluation frame, measuring in-context learning speed over 12–72 hour tasks and finding top frontier models doubling their learning rate every three months — a dimension established benchmarks do not capture[17].

Open-weights models reaching frontier-adjacent scores on established benchmarks offer developers cheaper, inspectable alternatives to proprietary APIs. The distillation question is unresolved: if GLM-5.2's scores partly reflect training on closed-model outputs, some of the apparent capability gain is an artifact. EdgeBench adds a more fundamental challenge — models that score well on one-shot tasks may still lag significantly on in-context learning, which is what agentic work actually requires.

Since late June 2026, four open-weights models have produced benchmark results closing 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[1] — leads open-weights models on the DeepSWE software engineering leaderboard[2] 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%[3]. 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[4]. Ornith-1.0-397B from DeepReinforce — 397B total parameters, 17B active — scores 82.4 on SWE-Bench Verified[5]; Simon Willison tested its 35B GGUF locally and found proficient agentic performance at 103 tokens per second under Apache 2.0[6]. Agents-A1 (35B) trains on 45K-token verified multi-step trajectories and distills specialist teacher models for search, science, and tool use, claiming 1T-parameter-model performance under Apache 2.0[7].

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[8], 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[9]. An independent investigation by Lisan al Gaib into GLM-5.2's post-training methods found the results "sus," amplifying rather than resolving the distillation concern[10]. Against this skepticism, hands-on evidence from multiple directions is consistently positive: bendee983, normally skeptical of benchmarks, found it genuinely impressive in direct use[11]; Semgrep reports it beating Claude in their internal cybersecurity benchmarks[12]; Julian Goldie SEO reports it winning 4 out of 5 real-world build comparisons against Claude Opus 4.8[13]; and developers report meaningful cost savings using it in Cursor[14]. A calibration analysis has used GLM-5.2's verified ARC-AGI-2 score to estimate closed model capabilities from eight private benchmark datapoints[15][16] — an approach whose reliability depends directly on whether GLM-5.2's scores are genuine or artifacts of distillation.

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 or static knowledge[17]. After 38,000 total agent-hours, performance as a function of environment interaction time follows a log-sigmoid curve. Top frontier models appear to be roughly doubling their 2-hour learning speed every three months — a potential scaling law for environment-based learning that most benchmarks used to rank GLM-5.2, Ornith-1.0, and VibeThinker-3B do not capture. ARC Prize has separately announced ARC-AGI-3 milestone prizes[18], indicating the abstract reasoning benchmark series is advancing past ARC-AGI-2.

The comparison baseline is also shifting on the closed-model side. Community commentary has introduced Fable 5 as a closed model that topped a physics benchmark at 6x the cost of Claude Opus 4.8[19][20], and Julian Goldie SEO has characterized Claude Sonnet 5 as scoring below Opus 4.8 on benchmarks despite higher pricing[21]. These observations, from social posts rather than formal evaluations, add a cost-per-benchmark-point dimension to the closed-model landscape that GLM-5.2's advocates cite when making the commercial case for open-weights alternatives.