Registry Documentation

About Benchmark Registry

Benchmark Registry is a public searchable database of artificial intelligence benchmark results. One benchmark record represents one reported metric result.

The registry combines results reported in provider reports, technical reports, system cards, papers, leaderboards, model cards, provider publications, and credible independent evaluations.

Understanding Record Numbers

A Benchmark Record Identifier is the stable public reference for one reported metric result. Examples include BR-00155-042, BR-002O48-017, and BR-00335F-006.

BR-00155-042
BR  = Benchmark Record
001 = OpenAI provider namespace
55  = Model Number 55, representing GPT-5.5
042 = result sequence 042 for this model-specific record prefix

Provider-specific encodings vary. Anthropic tier codes, for example, occur before the Model Number. There is no universal model format.

Understanding Model IDs

Model Identifiers are stable registry references. Examples: OPNAI-56 — GPT-5.6; ANTHR-O48 — Claude Opus 4.8; GOOGL-G35F — Gemini 3.5 Flash; META-MS — Muse Spark; and XAI-G45 — Grok 4.5.

Approved numeric Model Numbers omit the decimal point: 5.6 → 56, 4.8 → 48, and 3.5 → 35.

Identifier construction is provider-specific. Users do not need to understand Model Identifiers to search by official model name.

Provider identifier reference

Table may be scrolled horizontally on narrow displays.

Search Behavior

Search is case-insensitive. Recognized aliases may resolve to a canonical model.

Benchmark Versions and Variants

A benchmark family is the recognized benchmark identity. GPQA and SWE-bench are families; GPQA Diamond and SWE-bench Verified identify known variants. Searches may resolve a family across its versions or one exact known version or variant.

Two records using the same benchmark family may have used different benchmark variants or evaluation settings. A missing version remains explicitly unspecified; it is not represented as a fabricated numeric version.

Evaluation Context

Where sources support it, records preserve shots, reasoning mode, reasoning effort, pass count, agent scaffold, evaluation harness, sampling values, and other structured settings. Equivalent configurations share one stable internal reference through an exact canonical-data fingerprint.

Unknown means the registry does not currently have sufficient source-supported information. Unknown does not mean a default value was used.

Model Snapshots

A canonical model may be evaluated through a dated or provider-specific model snapshot. For example, canonical model OPNAI-55 may have an explicitly reported provider model identifier such as gpt-5.5-2026-06-18. Snapshot attribution is preserved only when reported; not all providers expose snapshot identifiers.

Comparing Benchmark Records

The home page can sort reported records by record number, model, benchmark, metric, stored numeric score, evaluation date, or Model Identifier. Score ordering is not a normalized model ranking: scores from different benchmarks or metrics may use different scales and meanings.

Benchmark Registry does not assume that records sharing a benchmark name are directly comparable. Benchmark variant, evaluation configuration, model snapshot, and evaluator may affect a reported result. Use the model, benchmark, and metric filters to narrow the table, then review each record's evaluation context before drawing comparisons.

Any structural compatibility assessment is conservative data-integrity support. It is not a scientific or statistical conclusion and does not rank models.

Sources

Benchmark results link to their underlying sources. Accepted categories include provider reports, system cards, technical reports, papers, official leaderboards, model cards, provider pages, and credible independent evaluations.

Source attribution and independent reproduction are different concepts. The registry does not claim to independently verify every reported result.

Record Statuses

ACTIVE

The record is currently part of the active registry.

WITHDRAWN

The record has been withdrawn but remains reserved and may remain searchable.

SUPERSEDED

A newer or corrected record has replaced the record.

ERRONEOUS

The record contains a known material error.

ARCHIVED

The record is retained for historical reference.

Identifiers are not reused.

Identifier Stability

Published Model Identifiers and Benchmark Record Identifiers are stored public references. Generation rules apply only when new identifiers are assigned. Existing identifiers are not recomputed when generation logic changes. Model renames do not automatically create a new identifier, and withdrawn identifiers are not reassigned.

Registry Disclaimer

Benchmark Registry stores benchmark results as reported by identified sources. Inclusion in the registry does not constitute endorsement of a model, provider, benchmark, evaluation methodology, or reported result. Unless explicitly stated, Benchmark Registry has not independently reproduced the evaluation.

Contact Us

To report a correction, include the stable Benchmark Record Identifier and a supporting source link.

X: @rolling7ho