A ranking is only as useful as the method behind it. The Fast Model Index exists to give U.S. professionals a living reference of fast AI models scored on the dimensions that actually matter in daily work. This article explains how the scores are constructed, what each dimension measures, how the weights were chosen, and why the raw data always remains visible alongside any summary number.
I built the scoring approach the same way I used to build product scorecards: make the criteria explicit, keep the evidence inspectable, and never let a single composite figure hide the trade-offs. Readers who disagree with the weights should still be able to use the underlying measurements.
The Five Dimensions
Every model that enters the Index is evaluated on five dimensions:
Speed (25%)
Cost (20%)
Output quality (25%)
Reliability (20%)
Workflow fit (10%)
These percentages are deliberate. Speed and output quality receive the largest shares because they most directly shape the daily experience of using a model. Cost and reliability follow closely because both determine whether a model can be sustained in real operations. Workflow fit is smaller yet still present because a model that scores well on abstract metrics but fails on the specific tasks teams actually run is less useful than the numbers alone suggest.
The weights are not claimed to be universal. They reflect the priorities of the primary audience described in the site plan—product managers, marketers, independent developers, and small-team operators. Readers who weight cost or reliability more heavily can still read the component scores and apply their own arithmetic.
How Each Dimension Is Measured
Speed draws from the standard AI Speed Index baseline and concurrent-load tests. First-token latency, sustained tokens per second, total completion time, and consistency across repeated runs are the primary inputs. Models that maintain tight variance under modest concurrency score higher than models that post attractive single-run numbers but drift under load.
Cost is calculated from observed token counts on the same tasks used for quality and workflow testing, multiplied by the public rates in effect on the test date. The resulting cost-per-task figures are normalized so that models can be compared even when their pricing structures differ. Hidden multipliers such as retries and context overhead are included when they appear in the test data.
Output quality is assessed on the practical dimensions already used in Versus and Real Workflows articles: instruction following, fidelity to source material, structural cleanliness, and the amount of human editing still required. Qualitative notes are translated into a consistent internal scale so that models tested at different times remain comparable.
Reliability combines failure rate, truncation rate, consistency of behavior across repeated runs, and observed recovery under light concurrent load. A model that rarely fails and recovers cleanly scores higher than a model whose average speed looks strong but whose error rate forces extra application logic.
Workflow fit reflects performance on the specific real-world tasks that appear in the Real Workflows category—support triage, marketing operations, document summarization, data extraction, and feedback synthesis. It is the dimension most tightly tied to the site’s practical focus and the smallest by weight so that it does not overpower the more general measurements.
Why Raw Data Stays Visible
Composite scores are convenient. They are also lossy. Two models can reach similar overall numbers by opposite routes—one leading on speed and trailing on cost, another doing the reverse. Without the component scores and the underlying measurements, readers cannot see those differences.
Every Index entry therefore links back to the raw speed series, the cost calculations, the quality notes, and the workflow results that produced the scores. The Index is a summary layer, not a replacement for the evidence. Readers who want to re-weight the dimensions or challenge a particular judgment have the material to do so.
This practice also protects the Index against silent drift. When a model version changes or a provider updates pricing, the affected components can be re-tested and the score updated with a clear record of what shifted.
How Models Enter and Update
A model is eligible for the Index once it has completed the core AI Speed Index test suite: baseline latency and throughput, concurrent-load behavior, at least one substantial cost analysis, quality evaluation on matched tasks, and results from one or more Real Workflows tests. Partial data produces a provisional entry with clear gaps rather than a full score.
Updates follow the same transparency rules. New test results replace older ones only when the model version, the test method, and the date are recorded. Historical scores remain accessible so that trends over time stay visible.
The Index is refreshed on a monthly cadence for rankings, with weekly data updates as new measurements arrive. The monthly snapshot is the public reference point; the underlying articles continue to accumulate evidence between snapshots.
What the Index Does Not Claim
A scored ranking still has clear limits. It does not claim:
That the five dimensions and their weights are the only reasonable way to evaluate models
That laboratory conditions free of ordinary network variation were achieved
That every existing fast model has already been tested
That future model versions will preserve their relative positions
It claims only that the models currently in the Index were measured under the documented AI Speed Index methods, scored on the stated dimensions, and published with the raw evidence still attached. That is enough to serve as a practical reference while remaining open to challenge and revision.
How the Index Is Presented
The Fast Model Index itself appears as a living reference page with the current ranked list, the component scores, and links to the supporting tests. Each major update is accompanied by a short article that notes what changed, which models moved, and which new measurements were added.
Individual model entries follow the familiar closing structure used across the site so that readers can move smoothly from the summary score to the detailed evidence:
Use it when…
Skip it when…
What we measured
What we still do not know
Fast is a feature. Reliable is the product. The Fast Model Index exists to keep both statements measurable across the models that small U.S. teams actually consider. The scores are a convenience. The evidence underneath them is the substance. AI Speed Index publishes both.
Letters
No letters yet — be the first to write.