LLM Benchmark Snapshot, September 22, 2026: A Frozen Week at the Top

LLM Benchmark Snapshot, September 22, 2026: A Frozen Week at the Top

Every week we photograph the major LLM leaderboards and keep the receipts at /benchmarks/history. This week's photo, September 22, is notable for what did not happen: across all five sources we track, not a single top-3 position changed hands.

The leaders, by source

The sources still disagree

The disagreement pattern we flag every week holds: Anthropic leads general-purpose text on three boards, OpenAI leads coding (Aider) and webdev (LMArena) decisively. If you pick a model by "the leaderboard," the answer still depends entirely on which leaderboard matches your workload.

The action was off the leaderboards

A frozen week at the top does not mean a quiet week. The most-discussed launches were not chat models at all: TypeSafe's Jev and the open-source, Apache 2.0 Laya from India's Convai Innovations opened up the "System One" decision-model category - typed answers with calibrated probabilities instead of generated text. None of the boards above measure that kind of model yet, which is itself a gap worth watching: the benchmark ecosystem measures writing and coding, while a growing slice of production AI is fast classification. Our comparison: Laya vs Jev.

Browse the full snapshot at /benchmarks/snapshot/2026-09-22, or the live board at /benchmarks.