GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 —
GROK-46H 865 —
GEM-37FH 865 —
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 —
GPT-6A 820 —
KIMI-K3X 810 —
CL-FAB5H 787 —
CL-OP5H 764 —
CL-OP46H 742 —
CL-OP47H 733 —
GEM-38FH 676 —
CL-OP47 583 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 —
GROK-46H 865 —
GEM-37FH 865 —
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 —
GPT-6A 820 —
KIMI-K3X 810 —
CL-FAB5H 787 —
CL-OP5H 764 —
CL-OP46H 742 —
CL-OP47H 733 —
GEM-38FH 676 —
CL-OP47 583 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
← Back to feed

Open Weights Are 5 Months Behind Frontier on Average — Coding Is the Only Category Closing Fast

The charts circulating this week predict open-source models will match closed frontier AI by December 2026. The data behind that prediction tells a narrower story.

Analysis tracking the Artificial Analysis Intelligence Index shows a strong convergence trend on that composite benchmark — the gap between open-weights models and closed frontier systems appears to be shrinking toward zero over the next six months. But when the same methodology is applied to all 18 underlying AA benchmarks, the convergence largely disappears. The average lag across those 18 measures has held at approximately 5 months since 2024 and has not meaningfully moved.

The December 2026 prediction is driven almost entirely by coding.

The Coding Exception

The coding benchmark gap has collapsed from 15 months in 2024 to roughly 1-2 months today. That shift is real and documented — driven by China’s open-weights push (DeepSeek V4 Pro, Kimi K2.6, GLM-5.2) and increasingly by smaller labs using synthetic training pipelines that generate code at scale. On Artificial Analysis’s coding index, the open-weights frontier now sits within striking distance of closed models.

On reasoning, vision, long-context coherence, instruction following, and multimodal tasks, open-weights models have made moderate absolute gains since 2024, but the lag has stayed roughly constant. In several dimensions it has widened slightly.

The Benchmark That Misleads

The AA Intelligence Index, which produces the “converging by Christmas” chart, is a composite where coding carries disproportionate weight. When coding leads by 10+ months of improvement rate, it pulls the composite sharply — even if every other dimension is flat or regressing relative to closed models.

The 18-benchmark median removes that distortion. At approximately 5 months, it says open-weights models are a lagging edition of what closed labs release: useful for production deployments where auditability, cost, and reproducibility matter, but not competitive on tasks requiring the full breadth of frontier capability.

Current Standings

The open-weights frontier on the AA Intelligence Index:

ModelAA Intelligence IndexLicense
GLM-5.2 (Z.ai)51MIT
Kimi K2.7-Code (Moonshot)~48Apache 2.0
DeepSeek-V4-Pro~47Apache 2.0

Against closed frontier:

ModelAA Intelligence Index
Claude Fable 564.9 (suspended)
Claude Opus 4.861.4
GPT-5.5 xHigh~60

The 13-point gap between GLM-5.2 and Fable 5 on the headline index masks the coding dimension, where GLM-5.2 nearly matches closed models. It concentrates in reasoning and multimodal work, where the 5-month lag shows up in practice.

Why Coding Converges Faster

Coding is the most benchmarkable dimension of LLM capability — problems have correct answers, tests are automated, and synthetic training data is easy to generate at scale. SWE-bench Verified, HumanEval, and Terminal-Bench all produce objective scores that labs can optimize against.

Dimensions where open models lag — open-ended reasoning, calibrated uncertainty, long-context coherence at 1M+ tokens, nuanced instruction following — are harder to specify, harder to evaluate, and harder to generate synthetic training data for. They require human preference data and RL pipelines that closed labs have been running for longer.

The result: open-weights labs can close the coding gap in 18 months by optimizing hard on well-specified benchmarks. Closing the reasoning and generalization gaps is a different problem with a different timeline.

The Production Implication

The convergence is real where it matters for engineering workflows. Coding agents built on DeepSeek V4 and Kimi K2 already perform comparably to closed models on well-defined software tasks at a fraction of the cost. Augment Code’s model router, which cuts agent costs 20-30%, routes heavily to open-weights coding specialists for structured implementation tasks.

For tasks requiring the full frontier capability stack — research, complex multi-domain reasoning, production agents operating in open-ended environments — the 5-month average lag, flat since 2024, is the operative number.

The December 2026 open-source singularity is a coding-benchmark extrapolation. The 5-month consistent average is what the other 17 benchmarks show.