How the model cards are made
Where each claim on the model cards comes from, what the source labels mean, how percentiles are computed, and how the data is kept current. Last updated 2026-10-26.
What this is
An index of what each model is reported to be good and bad at, collected from vendor pages, independent leaderboards and practitioner write-ups. Every claim links to its source and shows the date when there is one. It is a reference, not a recommendation: a model that tops one benchmark can still be the wrong choice for your task.
Where claims come from
Each claim carries one of four labels, shown next to its source link.
- Vendor's own site. A vendor reporting on its own model, on its own site or model page.
- Vendor claim, relayed. A number the vendor reported, read from someone else's page: a news article, a reseller, an aggregator. The number is the vendor's; the page is not. Where possible the vendor's own page is better evidence.
- Independent evaluator. A third party that ran the evaluation or maintains the leaderboard.
- Practitioner report. Blog posts, forum threads and social posts from people who used the model.
A claim without a source link is not published.
Percentiles
A benchmark is scored only when at least four tracked models have a result on it. A model's percentile on a benchmark is the share of tracked models it beats, with ties counted as half. A capability group (coding, agentic coding, knowledge and so on) shows the average of its benchmark percentiles. Percentiles compare tracked models with each other, so they move when models are added. When a model has several results for the same benchmark, the independent one is used.
Benchmarks differ in harness, settings and version, and vendors choose which results to publish. Read the benchmark name and source before comparing two numbers.
Tiers
Tier (fast, standard, deep, frontier) places a model within its vendor's lineup by blended price, weighted toward cached input as in a typical coding-agent session. It is available for vendors with a defined lineup; others show a dash.
Freshness
Practitioner reports older than 90 days are labelled. They are not removed, because an old report can still be the only one. Labels are worked out when the site is built, so a report can pass the 90-day mark a little before its page shows it. Rows with no date are labelled undated. Prices come from the same table gjalla uses to price sessions, or from the vendor's pricing page when a model is not in that table.
How it is updated
A weekly job searches for new models, new benchmark results and new reports. Its changes are proposed as a reviewed change, and a person approves them before they appear here. The last updated date above moves only when the data changes.
Gaps
Coverage is uneven. Some models have vendor claims but no independent results, or the reverse. A dash means we have no data, not a low score.
Use the data
models.json has every model, claim, score, source and date. llms.txt lists the pages. To report an error or a missing source, email support@gjalla.io.