Every pixel of distance is a real score gap.
This page explains exactly how a benchmark result becomes a position on a track: the units, how gaps are judged, how streaks are counted, where each number comes from, and what is still placeholder data.
Current status: illustrative data. Scores, histories and news on this build come from a deterministic synthetic provider modelled on real leaderboards. The methods below are the ones the live site applies; only the inputs change.
Distance is data
Gaps on every track, bar and chart are drawn in the race's own units. Nothing is stretched to make a race look closer or more exciting.
Weekly snapshots
The board is frozen every Monday. Streaks, records and history all come from those snapshots, so they never change after the fact.
Transparent blends
Composite races show their weights, and you can change them. Every race lists the benchmarks behind it.
Clear labels
Placeholder data, editorial opinion and measured results are always labelled as what they are.
From benchmark to track position
Each race measures one capability in its natural unit. Units never mix inside a single race; composites convert to a common 0 to 100 scale first.
- 01
Collect
Results from the public benchmarks and arenas listed under each race.
- 02
Keep native units
Accuracy stays a percentage, preference stays Elo, horizon stays hours.
- 03
Snapshot weekly
Every Monday the latest score for every model is frozen.
- 04
Rank and measure
Models are ordered and the gap to the leader and the car ahead is recorded.
- 05
Classify the lead
The gap is judged against the race's decisive gap to label the lead.
| Unit | What it measures | Decisive gap | Races |
|---|---|---|---|
| Score (%) | Share of problems solved, such as 86.4% | 12 | 35 |
| Elo rating | Rating from blind pairwise votes, such as 1,482 | 120 | 14 |
| Contest rating | Competitive programming rating, such as 2,950 | 500 | 1 |
| Task horizon (hours) | Length of task finished half the time, such as 16 hrs | 8 | 1 |
| Tokens per second | Output tokens per second, such as 180 tok/s | 150 | 1 |
| Seconds to first token | Time to first token, lower is better | 6 | 1 |
| Index (0-100) | Composite or cost index, 0 to 100 | 14 | 14 |
The decisive gap is the difference that separates tiers in that kind of measurement. Some races override the default; the value used is listed under each race below.
Photo finish or runaway lead?
A two-point lead means a lot on a hard benchmark and almost nothing on an easy one. So every lead is judged relative to that race's decisive gap, not in raw points.
Photo finishgap under 4% of decisive gap
Inside the noise of most benchmarks. Treat as a tie and pick on price, speed or fit.
Narrow leadgap 4 to 10% of decisive gap
Real but small. You may notice it on hard tasks; you will not on easy ones.
Clear leadgap 10 to 25% of decisive gap
Consistent and visible in day-to-day use.
Dominant leadgap 25% or more of decisive gap
A different tier. The leader does things the rest cannot do reliably.
On the 3D tracks, a full lap equals the decisive gap or the spread of the field, whichever is larger, so a dominant leader visibly laps the pack and a photo finish looks like one.
Try it: how big is a lead?
Gap
0.8 pts
Share of decisive gap
7%
Decisive gap: 12.0 pts
Status
Narrow lead
Weeks at #1, reigns and lead changes
Records only count weekly snapshots, so a model that briefly edged ahead between Mondays does not get credit.
W1
W2
W3
W4
W5
W6
W7
W8
W9
W10
W11
W12
W13
W14
Current streak
8 weeks
C has led every snapshot since week 7. Shown as "weeks at #1".
Total weeks at #1
A: 3
All snapshots a model led, streak or not. A's two separate reigns add up.
Longest reign
8 weeks
The longest unbroken run in the race's history. Feeds the reign records.
Lead changes
3
Every snapshot where the leader differs from the week before.
Streaks are always counted on the full field, including older versions, so they stay comparable when a lab ships a new model. See them all on the leaderboards.
Blended races you can re-weight
Headline races such as Overall and Coding blend their sub-races. Each sub-race is first mapped to 0 to 100, then combined with published weights.
Normalising. Percentages and Elo map linearly across each race's scale. Speed and time to first token use a log scale, because going from 20 to 40 tokens a second matters as much as going from 200 to 400. Task horizons use log base 2, so every doubling of task length is worth the same.
Weighting. Defaults favour the sub-races most people rely on, for example agentic and debugging work inside Coding. Drag the sliders on any composite race and every standing, gap, streak and chart recomputes with your weights.
Missing scores. If a model has no result in a sub-race, the remaining weights are rescaled. A model is only ranked in a composite when at least one sub-race has a score.
Latest versions. Races show the newest model in each family by default. History, records and head-to-head keep every version.
Composite races and their default weights
- Overall Intelligence10 headline races
- Coding11 sub-races
- Hard Reasoning4 sub-races
- Math3 sub-races
- Science4 sub-races
- Writing & Language6 sub-races
- Multimodal Understanding4 sub-races
- Tool Use & Agents4 sub-races
How agent task rankings work
Agent rankings score complete jobs, not single prompts: the agent gets a goal, tools and a budget, and is graded on whether the job is done.
Each task, such as fixing a failing test suite or reconciling an invoice batch, has a 0 to 100 index that blends success rate (most of the weight), cost per completed job and how often a human had to step in. The index is reported for a model with its recommended setup, so it reflects the model and the scaffold together.
Because every task shares the same 0 to 100 scale, leads are judged in fixed index points rather than a per-race decisive gap.
See the task rankingsNews, profiles and guides
Measured results and editorial judgement are kept apart, and each is labelled.
News impact scale
Only stories rated 3 or higher are published. Every story carries a plain “why it matters” and a provenance note.
5 · Landmark
Moves the frontier or changes how the field works: a new overall leader, a step change in capability, binding regulation.
4 · Major
Shifts a race or a product millions of people use, or meaningfully changes cost or access.
3 · Notable
Matters to a specific audience such as developers, creators or a single industry. The lowest level we publish.
Model profiles are written from lab documentation, independent evaluations and hands-on testing. Strengths and weaknesses are tied to race results where possible.
Guide digests in the prompting hub summarise each publisher's official guidance in our own words and link to the original.
No paid placement. Rankings, recommendations and news are never sponsored.
Where every race comes from
All 67 races, the benchmarks they are designed around and how each score is defined.
Open, static JSON
Everything on the board is available as static JSON, regenerated with each weekly snapshot. Use it for your own dashboards, with a link back.
- /api/races.jsonEvery race: leader, margin, status, streak and the top ten with gaps.
- /api/models.jsonModel directory with specs, pricing and primary rank.
- /api/leaderboards.jsonWeeks at #1, reign records and lab wins.
- /api/agents.jsonAgent task rankings with margins.
- /api/news.jsonCurated news with impact and why it matters.
- /api/prompts.jsonThe prompt library with variables and best-fit models.
Built to swap in live data
All content is typed, JSON-serialisable data: models, races, agents, prompts, lessons, quizzes and news. Scores flow through a single provider function, so connecting a benchmark database or CMS replaces one module while every race, chart, record and API endpoint keeps working.
scoreAt(model, category, weekIndex) -> number | null // native units