LuminaBench
Writing benchmark / Towards AI

ToneBench

Lumina VerifiedLumina checked the published sources, versions, model labels, score extraction and conversion. External evaluations were run by their publishers; Lumina has not independently rerun or reproduced them.

YouTube script writing in the voice of the What's AI channel.

Version Oct 2026 · 10 scriptsVerified 2 November 2026141 published resultsOfficial source

ToneBench · Lumina Score

LuminaBench
Show
20 of 84 models · best variant per model
03006009001,2001,160Claude Opus 5.5Max1,134Claude Sonnet 5.5Max1,134Claude Fable 5.1Max1,113GPT-6 SolMax1,108Claude Opus 5Max1,097Claude Fable 5Max1,097GPT-6.1 SolMax1,097Kimi K3.1 ProMax1,095GPT-5.6 SolProvider default (Low)1,095GLM-5.3Provider default (Max)1,093Kimi K3Provider default (Max)1,093GPT-6 AstraMax1,085GLM-5.3 FlashProvider default (Max)1,084DeepSeek V4.1 Flash 1017Max1,079GPT-5.6 TerraUltra1,076DeepSeek V4.1 FlashMax1,073Muse WonderMax1,072Claude Opus 4.8Max1,072Claude Haiku 5.5Max1,072MiMo-V2.6-Pro UltraSpeedthinking on

Lumina Score vs Cost per Script

API-equivalent list-price comparison; free routes remain zero

LuminaBench
15 of 84 families
$0$1$2$38009001,0001,1001,200Cost/script · USDLumina ScoreDeepSeek V4.1 Flash 1017GLM-5.3 FlashGPT-6 SolClaude Opus 5.5Claude Sonnet 5.5Claude Fable 5.1Claude Opus 5Claude Fable 5GPT-6 Astra

Lumina Score vs Output Tokens per Script

Publisher output counts, reasoning included where reported

LuminaBench
15 of 84 families
3K10K30K100K8009001,0001,1001,200Mean output tokens/script (log scale)Lumina ScoreClaude Opus 5.5Claude Sonnet 5.5Claude Fable 5.1GPT-6 SolClaude Opus 5GPT-5.6 SolClaude Fable 5GPT-6.1 SolKimi K3.1 ProGLM-5.3GPT-6 AstraGPT-5.6 TerraKimi K3

Lumina Score vs Time per Script

Mean accepted-script time, retries included

LuminaBench
15 of 84 families
13108009001,0001,1001,200Mean time/script · min (log scale)Lumina ScoreClaude Opus 5.5Claude Sonnet 5.5Claude Fable 5.1GPT-6 SolClaude Opus 5Claude Fable 5GPT-6.1 SolKimi K3.1 ProGPT-5.6 SolGLM-5.3GPT-6 AstraDeepSeek V4.1 Flash 1017Kimi K3

Writing Elo from pairwise comparisons of rubric scores on ten fixed scripts, five drafts each, judged by the current four-judge panel. The 0–100 Writing Score and rubric metrics are supporting evidence. Requested effort, provider default and thinking toggles stay distinct; intervals resample drafts with scripts fixed. Missing metrics are omitted from their graph; quality scores remain available.

Results table

Method & coverage

Writing Elo from pairwise comparisons of rubric scores on ten fixed scripts, five drafts each, judged by the current four-judge panel. The 0–100 Writing Score and rubric metrics are supporting evidence. Requested effort, provider default and thinking toggles stay distinct; intervals resample drafts with scripts fixed.

Publisher mean API-equivalent USD/script, output tokens/script (reasoning included where counted) and mean seconds/script including retries. List-price comparison, not actual subscription spend; free routes retain zero. Cost method and estimation flags preserved.

How the Lumina conversion works

1,000 + 75 × ((original score − 1317.9051587301587) / 676.1206043288362). The result is rounded once. Scores may tie.

Frozen initial eligible publisher cohort; one arithmetic mean per exact release/system, equally weighted across releases. Subsequent refreshes never refit. Strictly increasing conversion preserves source ordering at full precision; rounded display scores may tie.

Calibration
tonebench-z75-v1 · experimental
Reference mean
1317.9051587301587
Population SD
676.1206043288362
Coverage
Standalone benchmark.
Identity, exclusions & attribution

Exact source release, effort, route and harness retained. Unknown evaluation dates stay unknown; no cross-source score estimates or combined indexes.

Code licensing is not assumed to clear separately published result redistribution.

All published reasoning settings, including Claude Max, are included by default. Display labels use Low and xHigh; source labels remain in the retained records.

Official results

Publisher terms. Scores and resource figures are attributed to the original publisher.