Where do these scores come from?

Every score comes from a public benchmark or official machine-readable file, with a source link and date. This site does not invent scores, and missing values are not treated as zero.

How is the overall score calculated?

Each source is min-max normalized within itself, then weighted: Artificial Analysis 28%, LMArena Text Arena 28%, LiveBench 22%, coding composite 15%, reasoning composite 7%. Missing sources are dropped and the rest are renormalized. A model needs at least two headline sources (or three partials) to enter the overall board. When several variants of one family exist, only the best overall variant is shown.

Are site prices part of the benchmarks?

No. Benchmark scores come from third-party public boards. CNY prices are this site’s catalogue reference rates (CNY 1 = 1,000 points) and are settled on actual usage. They are shown separately so billing is not confused with quality.

How often is this updated?

A collector runs automatically every day. The upstream boards may not publish a new snapshot daily, so each page shows both this site’s collection time and the source publication date.

88Z.ai / RANKINGS

Leaderboard methodology

The rules are written out so people, search engines and language models can cite them.

Last collected by this site: 2026-11-15 06:40 (UTC+8, Asia/Shanghai). Source publication dates are in the table or the citations below.

Principles

Only publicly verifiable evaluation data is used. Every score has a source URL and a date.
This site does not rerun the benchmarks, and it does not write missing values as zero.
The overall board is family-based: when several variants exist, the best overall variant is shown and labelled.

Overall weights

Component	Weight	Source metrics
Artificial Analysis Intelligence Index	28%	aa_intelligence
LMArena Text Arena	28%	lmarena_text_overall
LiveBench 2.0 overall	22%	livebench_overall
Coding composite	15%	coding_composite
Reasoning composite	7%	reasoning_composite

Each metric is min-max normalized to 0–100 on the current sample, then weighted as above. If a model is missing a component, the remaining weights are renormalized. To enter the overall board, a model needs at least two of Artificial Analysis, LMArena Text and LiveBench, or at least three component scores.

Category boards

The coding composite averages whichever of WebDev, LiveBench coding, Aider Polyglot and SWE-bench Verified (mini-SWE-agent protocol) are available, after normalization. The reasoning composite averages LiveBench reasoning, LMArena math and SuperCLUE math reasoning. The Chinese board leads with SuperCLUE totals and also shows LMArena Chinese battle scores. Vision, image and video boards use the corresponding Arena overall scores. Value combines quality and cost; speed uses Artificial Analysis time per task. Popularity uses Text Arena battle counts, not traffic on this site.

What should not be mixed casually

Benchmarks differ in samples, prompts, tool access and grading. Arena scores are relative battle ratings, not percentage accuracy. For SWE-bench this site only uses the mini-SWE-agent protocol, so results from other agent scaffolds are not treated as a model-to-model gap. Open-source labels follow the license field in the source and may be incomplete.

Sources

LMArena Text Arena (style-controlled) — Official Arena leaderboard dataset, text_style_control / latest. Scores are Arena ratings (Bradley-Terry).
LMArena WebDev Arena — Official webdev / latest snapshot, measuring front-end and web-app battles.
LMArena Vision Arena (style-controlled) — Official vision_style_control / latest snapshot.
LMArena Text-to-Image Arena — Official text_to_image / latest snapshot.
LMArena Text-to-Video Arena — Official text_to_video / latest snapshot.
LMArena Agent Arena — Official agent / latest aggregate scores.
Artificial Analysis Intelligence Index — Intelligence Index, cost per task, time per task and first-token latency from the official public leaderboard page.
LiveBench 2.0 — Official CSV results. Overall uses the published formula: Long Agentic ×2, Design ×1.5, Coding ×1.5, other suites ×1, then a weighted mean over suites that have a score.
Aider Polyglot — Official polyglot_leaderboard.yml. The metric is pass_rate_2 on 225 cross-language exercises. Repeated submissions keep the best score.
SWE-bench Verified (mini-SWE-agent) — Official leaderboards.json, Verified board, resolved % under the mini-SWE-agent protocol so models are compared in a similar scaffold. Scores from other agents are not mixed into the composite coding score.
SuperCLUE Chinese general board — Total and section scores from the official monthly general-board spreadsheet.

FAQ

Where do these scores come from?: Every score comes from a public benchmark or official machine-readable file, with a source link and date. This site does not invent scores, and missing values are not treated as zero.
How is the overall score calculated?: Each source is min-max normalized within itself, then weighted: Artificial Analysis 28%, LMArena Text Arena 28%, LiveBench 22%, coding composite 15%, reasoning composite 7%. Missing sources are dropped and the rest are renormalized. A model needs at least two headline sources (or three partials) to enter the overall board. When several variants of one family exist, only the best overall variant is shown.
Are site prices part of the benchmarks?: No. Benchmark scores come from third-party public boards. CNY prices are this site’s catalogue reference rates (CNY 1 = 1,000 points) and are settled on actual usage. They are shown separately so billing is not confused with quality.
How often is this updated?: A collector runs automatically every day. The upstream boards may not publish a new snapshot daily, so each page shows both this site’s collection time and the source publication date.