Token efficiency: quality per dollar of real work
BenchLeader Index divided by what an average task of Artificial Analysis' Intelligence Index cost on that configuration. Unlike price per token this counts the tokens a model actually burns, so a model that thinks at length scores lower than its sticker price suggests.
As of 15 Nov 2026, BenchLeader ranks GPT-6 Luna first on the Token efficiency: quality per dollar of real work board at 8945, ahead of Claude Haiku 5.5 (1090) and Granite 4.2 8B (1016).
120 models · data as of 15 Nov 2026
Quality against what it costs to get there
Hover or tap a model to pick it out and see its other effort settings; click to pin it (Esc clears). The eye hides a model.
Cost per task is what an average task of the Intelligence Index cost at list price, averaged over Artificial Analysis’ whole run of that configuration (source: Artificial Analysis). It counts the tokens a model really writes, so a model that thinks at length pays for it here. Each dot is a model at its best reasoning-effort setting (highest Index); pick a model out to see its other settings, joined lowest effort first; up and to the left is better.
How far the effort dial moves each model
One bar per model, from its lowest to its highest figure across its reasoning-effort settings, a mark at each setting. The same models as the chart above.
Tokens per task
output tokens, reasoning includedOutput tokens, full Index run
every task of the run, as AA totals itArtificial Analysis’ own total for the whole run: nine evaluations and over 5,000 tasks, several run three or five times. Its per-task figures are averages weighted by each evaluation’s share of the index, so they do not multiply back to this.
Cost per task
USD at list priceCost of a full Index run
USD at list price, as AA totals itArtificial Analysis’ own total for the whole run: nine evaluations and over 5,000 tasks, several run three or five times. Its per-task figures are averages weighted by each evaluation’s share of the index, so they do not multiply back to this.
Time per task
per Index taskCost, tokens and time per task, and the full-run totals, are from Artificial Analysis’ Intelligence Index runs, shown with credit and not offered for reuse. The Index is BenchLeader’s.
- 1GPT-6 Luna (low)8945
- 2Claude Haiku 5.5 (low)1090
- 3Granite 4.2 8B1016
- 4Mistral Small 4 (thinking)1015
- 5GPT-6.1 Terra (low)793
- 6MiMo-V2.6-Flash426
- 7Muse Ember (xhigh)422
- 8GPT-6.1 Sol (low)372
- 9Gemini 3.1 Flash Lite270
- 10Muse Glimmer (high)261
- 11Inkling Small256
- 12MiMo-V2.6-Pro221
- 13GPT-5.6 Luna (max)170
- 14Beam 1.0161
- 15Nemotron 3.5 Lightning132
Cite as: BenchLeader, “Token efficiency: quality per dollar of real work”, data as of 15 Nov 2026, https://www.benchleader.com/leaderboards/efficiency.