Blog›Guides

LLM Benchmark Snapshot: August 23, 2026

The first weekly photo of the LLM leaderboards: LMArena, LiveBench, BenchLM and Aider, frozen on one citable page. Anthropic sweeps three podiums; GPT-5 keeps the coding crown.

LLM Benchmark Snapshot: August 23, 2026

This is the first entry in a new Wikiprompt series: every week we freeze a complete photo of the major LLM leaderboards and publish it as a permanent, citable page. The full tables for this snapshot live at /benchmarks/snapshot/2026-08-23, and every future photo will be indexed at /benchmarks/history.

Why dated snapshots?

Benchmark numbers move quietly: a model climbs, a score is recalculated, and last month's leader silently becomes fourth. By keeping one URL per weekly photo, the record stays public. You can cite "the rankings as of August 23, 2026" and that link will always show exactly that.

What the photo shows

This snapshot covers five independent, open sources, each with its own methodology:

LMArena (text arena): claude-opus-5-high leads the community Elo, with claude-opus-5-max and claude-opus-4-6-high right behind. Anthropic holds the whole text podium this week.

LMArena (web development): claude-opus-5-max on top, followed by kimi-k3-max and qwen3.8-max, two open-weight challengers from Moonshot and Alibaba pressing the frontier labs.

LiveBench: claude-fable-5-max-effort takes first with an average of 83.4, ahead of gpt-5.6-sol-max (81.7) and gpt-5.5-xhigh (80.8).

BenchLM aggregate: an all-Anthropic podium: Claude Mythos 5 (83.0), Claude Opus 5 (82.7) and Claude Fable 5 (82.7), separated by a quarter of a point.

Aider polyglot (coding): gpt-5 (high) still rules practical coding with an 88.0 percent pass rate, ahead of gpt-5 (medium) and o3-pro (high).

The interesting read is the disagreement: community voting (LMArena), academic-style task suites (LiveBench), aggregated scores (BenchLM) and hands-on coding (Aider) do not crown the same model. That disagreement is exactly why we track many sources instead of one.