Benchmarks

Benchmark boards compare models under a particular evaluation, metric and scope. Entries retain source-native scores, ranks, uncertainty and sample sizes where supplied. The board directory and source filters help you select comparable results.

Public source pages use /benchmarks/{source_id}. Each indexed benchmark has a page at /benchmarks/{source_id}/{benchmark_id}; ?leaderboard={leaderboard_id} opens a particular evaluation cohort. These pages consume the same versioned rankings API and preserve separate metric and provider conditions. Model-card scores link to their indexed cohort; a score without indexed identifiers keeps its verified official source link. Inference-provider labels open the internal provider catalog. Official data remains linked within each benchmark page.

/leaderboards pages boards and caps the number of entries within each board. /insights provides curated chart cohorts with their source-native comparison context. Model scores list observations already linked to a canonical identity. For complete structured observations, including unresolved models and individual runs, use Evaluations.

Model names use the same destination rule in every benchmark surface: a non-null canonical_model_id opens the indexed TensorTrace model page; otherwise an optional model_url opens the model URL reported by the original result or official publisher metadata. If neither is supplied, render plain text. Do not invent a catalog search link or use the benchmark's source_url as a model URL. The latter remains a separate evidence link. A canonical association without a prepared model page is not advertised as an internal destination.

Arena's official parquet export does not include model URLs. Scheduled ingestion supplements it with exact-name model-link metadata from the published JSON on the corresponding public leaderboard. Successful metadata reads are reused for 24 hours; unavailable optional metadata preserves previously recorded links. Scores still come from the parquet export. The model-link source page, retrieval time and checksummed original published JSON remain in source evidence.

Never combine ranks from different boards into a single claimed ranking. Units, direction, versions, reasoning settings and workload conditions can change what a score means.

The public benchmark directory and cohort pages use the same chart presentation. Each chart links to official data through the reported source page or entry source_url; the latter falls back to the registered source website. Retired observation-list URLs redirect to the chart directory, retaining source and benchmark filters and resetting offsets. Raw records, unmatched identities and non-comparable results continue to be collected and remain available through /api/v1/evaluations; they are not turned into invented chart scores.

OpenASR refreshes use only the officially advertised, public table callbacks on its Hugging Face Space. English, seven multilingual language views and long-form results have separate metric cohorts; selected datasets, publisher version, original ranks, queries and checksummed raw responses remain in source evidence. WER is lower-is-better and RTFx is higher-is-better. Private, VoiceArena, benchmark-fitting and Hindi views are excluded. Public readability does not establish a result license; unknown licenses remain unknown.

The current public views replace only openasr-public-v1: observations. Earlier CSV acquisitions retain their original records, receipts, aliases and metric cohorts. They are not withdrawn when a smaller public view is published, nor combined into an average with that view. Missing views or changed columns fail the acquisition before publication and preserve the last successful results.

On this page

List ranked benchmark boards

GET /api/v1/leaderboards

Scores grouped by evaluation; ranks retain the upstream ranking scope.

Access: public read; no TensorTrace API key.

Parameters

NameLocationRequiredType / limitsMeaning
qqueryNostringCase-insensitive search text. Omit it to browse the complete matching catalog.
sourcequeryNostringRegistered data-source ID, such as openRouter. This identifies the collecting source, not necessarily the model developer.
dimensionqueryNostringExact benchmark comparison dimension; source-native dimensions are available in /facets/leaderboards.
benchmarkqueryNostringCase-insensitive substring of the benchmark name. It does not merge different benchmark versions or conditions.
benchmark_idqueryNostringExact source-native benchmark ID. Prefer it when a name substring would match multiple benchmarks.
leaderboardqueryNostringCase-insensitive substring of the comparison-board name.
leaderboard_idqueryNostringExact comparison-board ID; ranks are meaningful only within the same board and conditions.
entry_limitqueryNointeger Minimum: 1; Maximum: 100Maximum number of ranked entries returned within each board. Default 50; clamped to 1–100.
limitqueryNointeger Default: 500; Minimum: 1; Maximum: 500Maximum number of returned rows. The documented default and range apply; numeric pagination limits are clamped unless stated otherwise.
offsetqueryNointeger Default: 0; Minimum: 0Zero-based number of matching rows to skip. Default 0; negative values are clamped to 0.
entry_offsetqueryNointeger Default: 0; Minimum: 0; Maximum: 1000000Offset within each ranking cohort; preserves published ranks.

Request example

curl 'https://tensortrace.net/api/v1/leaderboards?limit=1'

Responses

StatusMeaningBody type
200Source-backed JSON responseLeaderboardsResponse
400Invalid queryNo declared JSON body
404Unknown model/sourceNo declared JSON body
500Database unavailableNo declared JSON body
503Service temporarily unavailableNo declared JSON body

Response example · 200

Illustrative values; schema-valid and not a live catalog measurement.

{
    "data": [
        {
            "benchmark_id": "example",
            "leaderboard_id": "example",
            "benchmark_name": "example",
            "dimension": "example",
            "entries": [
                {
                    "source_id": "openRouter",
                    "canonical_model_id": "google/gemini-2.5-flash",
                    "display_name": "Model Reviewer",
                    "provider": "google",
                    "unit": "input_tokens_per_million",
                    "source_record_id": "google/gemini-2.5-flash"
                }
            ],
            "source_id": "openRouter",
            "metric": "example",
            "unit": "input_tokens_per_million"
        }
    ],
    "total_count": 1
}
FieldTypeRequiredMeaning / constraints
dataarray of LeaderboardDtoYes
total_countintegerYes

List model benchmark scores

GET /api/v1/models/{provider}/{model}/scores

Access: public read; no TensorTrace API key.

Parameters

NameLocationRequiredType / limitsMeaning
providerpathYesstringCanonical model publisher namespace; URL-encode this path segment.
modelpathYesstringCanonical model name within the publisher namespace. URL-encode nested slashes in this single segment.
sourcequeryNostringRegistered data-source ID, such as openRouter. This identifies the collecting source, not necessarily the model developer.
limitqueryNointeger Default: 50; Minimum: 1; Maximum: 100Maximum number of returned rows. The documented default and range apply; numeric pagination limits are clamped unless stated otherwise.
offsetqueryNointeger Default: 0; Minimum: 0Zero-based number of matching rows to skip. Default 0; negative values are clamped to 0.

Request example

curl 'https://tensortrace.net/api/v1/models/google/gemini-2.5-flash/scores?limit=1'

Responses

StatusMeaningBody type
200Source-backed JSON responsearray of BenchmarkScoreDto
400Invalid queryNo declared JSON body
404Unknown model/sourceNo declared JSON body
500Database unavailableNo declared JSON body
503Service temporarily unavailableNo declared JSON body

Response example · 200

Illustrative values; schema-valid and not a live catalog measurement.

[
    {
        "source_id": "artificialAnalysis",
        "benchmark_id": "intelligence-index",
        "benchmark_name": "Intelligence Index",
        "leaderboard_id": "example",
        "dimension": "capabilities",
        "metric": "intelligence_index",
        "unit": "score",
        "score": 56,
        "source_record_id": "google/gemini-2.5-flash",
        "source_model_id": "google/gemini-2.5-flash"
    }
]

Type: array of BenchmarkScoreDto.

Curated source-native chart cohorts

GET /api/v1/insights

Latest stored receipt per source. Eleven Arena overall task cohorts (top 12 each), BFCL overall (up to 500), and SWE-bench Verified systems (top 12). Includes unmatched source models. Percent scores and bounds use 0–100 units; Arena net improvement can be negative. Counts cover the complete scored cohort before response truncation. Publication and retrieval dates are distinct; missing publication dates are null. Uses existing scheduled source refreshes. No cross-source composite score, historical series, or Arena task-cost data is implied. Model destinations prefer non-null canonical_model_id (a prepared TensorTrace model page), then the optional model_url preserved from original result or official publisher metadata. With neither, render a plain name. source_url remains the benchmark/export evidence, not a fallback model destination.

Access: public read; no TensorTrace API key.

Parameters

This operation has no path or query parameters.

Request example

curl 'https://tensortrace.net/api/v1/insights'

Responses

StatusMeaningBody type
200Source-native chart data; empty boards remain explicit.InsightsResponse

Response example · 200

Illustrative values; schema-valid and not a live catalog measurement.

{
    "boards": [
        {
            "id": "google/gemini-2.5-flash",
            "title": "example",
            "source_id": "openRouter",
            "source_url": "https://example.com/model",
            "metric": "example",
            "unit": "rating",
            "methodology": "example",
            "published_at": "2026-10-05",
            "retrieved_at": "2026-10-05T00:00:00Z",
            "receipt_id": "00000000-0000-4000-8000-000000000001",
            "source_record_count": 1,
            "scored_entry_count": 1,
            "linked_entry_count": 1,
            "entries": [
                {
                    "source_record_id": "google/gemini-2.5-flash",
                    "canonical_model_id": "google/gemini-2.5-flash",
                    "label": "example",
                    "organization": "example",
                    "score": 1,
                    "rank": 1,
                    "confidence_lower": 1,
                    "confidence_upper": 1,
                    "sample_size": 1,
                    "sample_unit": "example",
                    "cost_usd": 1,
                    "latency_seconds": 1,
                    "profile": "example",
                    "submitted_at": "2026-10-05",
                    "checked": false,
                    "warning": "example"
                }
            ]
        }
    ]
}
FieldTypeRequiredMeaning / constraints
boardsarray of InsightBoardYesMax. items: 13

Read report charts and real-world safety context inside Benchmarks

GET /api/v1/benchmarks/context

Reviewed Stanford charts, licensed AIID incidents and original AILuminate grades remain separate observations. Unknown denominators prohibit company-risk rankings.

Access: public read; no TensorTrace API key.

Parameters

NameLocationRequiredType / limitsMeaning
sourcequeryNostring Values: stanfordAiIndex, aiIncidents, ailuminateResults, mlCommonsOptional source; mlCommons is an alias for ailuminateResults.
kindqueryNostring Values: report_chart, incident, safety_gradeOptional observation domain; these are not interchangeable model scores.
groupqueryNostring Max. length: 1000Exact context_id of a reviewed report chart or incident.
limitqueryNointeger Default: 12; Minimum: 1; Maximum: 18Maximum number of context entries in one page.
offsetqueryNointeger Default: 0; Minimum: 0; Maximum: 2000000Number of matching entries to skip.

Request example

curl 'https://tensortrace.net/api/v1/benchmarks/context?limit=1'

Responses

StatusMeaningBody type
200Stored typed context with source/date coverage.object
400Invalid filters or temporarily unavailable database.ErrorResponse
503Invalid filters or temporarily unavailable database.ErrorResponse

Response example · 200

Illustrative values; schema-valid and not a live catalog measurement.

{
    "data": [
        {
            "source_id": "openRouter",
            "context_id": "example",
            "kind": "report_chart",
            "title": "example",
            "source_url": "https://example.com/model",
            "receipt_id": "00000000-0000-4000-8000-000000000001",
            "updated_at": "2026-10-05T00:00:00Z",
            "context": {
                "report_year": 1,
                "chart": {
                    "type": "bar",
                    "metric": "example",
                    "unit": "input_tokens_per_million",
                    "label_column": 1,
                    "value_column": 1
                },
                "columns": [
                    "example"
                ],
                "rows": [
                    [
                        "example"
                    ]
                ],
                "source_file": "https://example.com/model",
                "scope": "example"
            }
        }
    ],
    "total_count": 1,
    "canonical_source": "example",
    "summary": {
        "sources": [
            {
                "source_id": "openRouter",
                "count": 1
            }
        ],
        "evaluations": [
            {
                "source_id": "openRouter",
                "count": 1
            }
        ],
        "incident_years": [
            {
                "label": "example",
                "count": 1
            }
        ],
        "incident_risks": [
            {
                "label": "example",
                "count": 1
            }
        ],
        "incident_date_coverage": {
            "total": 1,
            "estimated_or_unknown": 1
        },
        "denominator": null,
        "company_risk_ranking": false
    }
}
FieldTypeRequiredMeaning / constraints
dataarray of BenchmarkContextYesMax. items: 18
total_countintegerYesMinimum: 0
canonical_sourcestring,nullYes
summaryBenchmarkContextSummaryYes

Models

Loading
EVIDENCE / COMPARISON

Benchmarks

Explore every benchmark

Loading

These charts contain model-linked results. The charts above also include unlinked source profiles. Compare scores within each evaluation. Model quality, usage, transparency and system performance measure different things.

Install TensorTrace