Methodology
TrustTheBench collects published measurements and shows where each one came from. We don't run our own model evaluations, and we never estimate a number we couldn't source. If a value is missing, the page says so.
Model data last refreshed 15 November 2026. Audio data last refreshed 15 November 2026. The dataset currently covers 499 models, 154 labs and providers, 92 benchmarks and 2,773 sourced scores.
Sources
| What | Source |
|---|---|
| Intelligence Index, per-evaluation results, cost and time to run, output speed, latency | Artificial Analysis |
| Release dates for tracked models | Artificial Analysis release tracker |
| Text-to-speech Elo (blind votes) | Artificial Analysis Speech Arena |
| Speech-to-text word error rate, speed factor, price | Artificial Analysis speech-to-text |
| API prices, context length, modalities, Hugging Face ids | OpenRouter models API |
| Weekly token volume and week-over-week change | OpenRouter rankings |
| Benchmark papers, human baselines, contamination notes, historical scores | Peer-reviewed papers and public leaderboards (cited per benchmark) |
| Vendor facts that independent sources don't cover (e.g. Sarvam Bulbul, Saaras) | Vendor documentation and changelogs (cited per model) |
The no-estimates rule
A number appears only if a named source published it. We don't fill gaps with averages, derive one benchmark from another, guess prices from model size, or assign a human baseline that no paper reported. Benchmarks without a published human baseline say “Not established”. Models that only appear because a paper listed a run configuration (for example a few-shot variant) keep their sourced scores but are left out of model lists and counts.
Release dates
Each date is resolved in this order, and the badge beside it says which rule applied:
- Verified: Matches an official announcement we have on file. (108 models)
- Artificial Analysis: Release date from the Artificial Analysis release tracker. (195 models)
- Version date: Taken from the dated version id the model is served under on OpenRouter. (118 models)
- Listing date: First date the model appeared on OpenRouter; the true release may be earlier. (64 models)
- Unknown: No reliable release date is available. (14 models)
Matching is exact. To correct a date, the official announcement is added to the verified list with its source.
Headline evaluations
The leaderboard, charts and category pages summarise models on six axes. Each axis is one named, independently measured evaluation, chosen because current frontier models are reported on it:
- Reasoning: CritPt v2 (Hard)
- Coding & agents: Terminal-Bench 4.0
- Math: AIME 2025
- Long context: AA-LCR
- Knowledge: Humanity's Last Exam
- Instruction following: IFBench
“Headline avg” is the plain mean of the axes a model has results for, and needs at least three. It's a convenience summary, not a new benchmark; the leaderboard sorts by the Artificial Analysis Intelligence Index by default.
Audio
Speech Arena (Text-to-Speech) (Elo rating (higher is better)). Models are ranked with an Elo system derived from blind user votes: listeners hear two speech samples generated from the same text and pick the one that sounds more natural. The provider-voice leaderboard uses each provider's own native voices. Each result is reported with a 95% interval and the number of votes (samples).
AA-WER v2 (Speech-to-Text) (Word error rate, % of words transcribed incorrectly (lower is better)). Artificial Analysis Word Error Rate index for non-streaming transcription. AA-WER v2 combines three datasets: AA-AgentTalk (50%), VoxPopuli-Cleaned-AA (25%) and Earnings22-Cleaned-AA (25%). Measured independently on dedicated hardware.
Speed Factor (Speech-to-Text) (Input audio seconds transcribed per second (higher is better)). Median speed factor: how many seconds of input audio are transcribed per second of wall-clock time. Measured independently by Artificial Analysis.
Models that no independent leaderboard ranks yet (for example Sarvam's Bulbul and Saaras) are listed as vendor-reported, with specifications and sources but no scores. See Audio benchmarks.
Questions
Does TrustTheBench estimate or interpolate any numbers?
No. Every number is either published by a named source or left blank. Missing scores show a dash, and missing prices or dates are not filled in from similar models.
Why does a model show no Intelligence Index?
Artificial Analysis only runs its index on the models it tracks. Older or niche models still have prices, context and release dates, but no index until an independent measurement exists.
How are release dates decided?
In order: an official announcement on file, the Artificial Analysis release tracker, the dated version id the model is served under, or the day it first appeared on OpenRouter. The badge next to each date says which one applies.
Which benchmarks are still worth trusting?
Benchmarks that are not saturated and have low contamination risk, such as Humanity's Last Exam, CritPt, Terminal-Bench and AA-LCR. Saturated suites like MMLU, GSM8K and HumanEval no longer separate frontier models. Each benchmark page shows its status and the evidence for it.
How do I cite TrustTheBench?
Cite the original source shown next to the number (for example Artificial Analysis or the benchmark paper), and optionally TrustTheBench (https://trustthebench.com) with the date you accessed it.