AI Battle > About AI Battle

About AI Battle

What AI Battle is

AI Battle is an aggregator, not a benchmark lab. We collect public AI benchmark datasets and the list prices that AI providers publish, and we turn them into comparable, linked leaderboards. Every ranking on this site is a read-only view over one named source dataset: the page states which dataset it comes from, how many entries the source lists, and when the page was last rebuilt. If a number appears here, it can be traced back to the source it came from — we publish no metric we cannot point at.

What we aggregate

At the time of writing the corpus covers 129 datasets in 26 families:

Language models — intelligence indexes, response latency (end-to-end, time-to-first-token, reasoning time), throughput, and published API pricing including cache-hit and cache-write rates.
Agentic evaluation suites — coding agents, chatbots, and dedicated eval families: HLE, MMMU, terminal-bench style tasks, omniscience, LCR, openness, GDPval, τ³ and SciCode-style benchmarks.
Multimodal generation — text-to-image and image-editing arena Elo, image API pricing, video and image-to-video arena rankings, music (vocals and instrumental).
Voice — text-to-speech (characters per second, price, arena Elo), speech-to-text (streaming and batch), and speech-to-speech families.

Each family keeps its own entity type: models, gateways, agent products or generated-media systems. Entries carry the creator label the source publishes, where the source publishes one.

How the pipeline works

Every deploy re-pulls each dataset's rows from the upstream API. We keep the ordering the source publishes — AI Battle does not re-rank, re-weight or normalise scores. On top of the raw rows we compute, and clearly label as our own:

distribution statistics per ranking: median, mean, minimum, maximum, spread and the gap between first and second place;
a creator breakdown: how many entries each creator holds in the window and the highest published value among them;
a cross-dataset profile: where the same model appears in other AI Battle datasets and how highly it places there, joined by the model's slug;
a machine-readable copy of every ranking at /data/rankings/<slug>.json with the same rows, statistics and provenance.

What we do not do

We run no evaluations of our own. Every score is the source dataset's.
We take no payment for placement. Rankings are ordered by the source, and the source is named on every page.
We do not invent units or directions. The upstream API does not declare units or which direction is "better", so values are shown exactly as published; the delta column on a ranking page is a plain arithmetic difference against the first entry, not a judgement.
We do not average across benchmarks. The cross-dataset profile lists placements; it deliberately does not compute a single "overall score".

How to read a ranking

Each ranking page shows the complete set of entries the source returns — the page states "the source lists N entries and all N are shown here". The ordering is the source's own. Whether a higher value is better depends on the dataset: response times and prices are costs, intelligence and Elo scores are capabilities. The delta column compares each entry to the leader arithmetically; read it against the metric name.

About prices

Prices on AI Battle are list prices published by the providers, per the unit the dataset states (typically per million tokens, per thousand images or per minute). They exclude volume discounts, negotiated rates, batching, caching behaviour under your traffic pattern, and taxes. Providers change prices without notice — before any purchasing decision, verify the price on the provider's own pricing page. The machine-readable price table lives at /pricing.md.

Limitations

A benchmark score describes one evaluation run by the source under its own conditions; it is not a prediction of quality on your workload.
Arena-style Elo rankings move as new votes arrive; a position is a snapshot.
Creator labels are reproduced as the source spells them; the same company can appear under more than one label, and the creator breakdown does not merge them.
Datasets differ in coverage: the same model can appear in one dataset and not in another, which is why the cross-dataset profile reports placement per dataset instead of pretending to a total.
Where a source publishes no per-entry creator label, the ranking says so rather than guessing.

Corrections

If a value here is wrong, the first fix is upstream: each ranking page names its dataset, and we link the source so you can check it directly. If the error is in our aggregation, statistics or labelling, write to hi@zordonintelligence.pl — we correct or annotate the affected pages, and this page records what changed.

Who runs AI Battle

AI Battle is built and operated by Zordon Intelligence sp. z o.o., a company registered in Poland. Data handling is described in the privacy policy: no account is required, request telemetry is anonymised, IP addresses are truncated, and EEA/UK/Swiss visitors get consent signals under IAB TCF v2.3 before any advertising cookies load.

Methodology changelog

2026-09-18 — ranking pages gained computed distribution statistics, creator breakdowns, cross-dataset profiles and per-ranking JSON data files; every page now states its window completeness explicitly.