Towards AITowards AIToneBench

Lab Comparison

Anthropic Models Compared

Last updated , from board data generated at 17:52 UTC.

All 34 ranked Anthropic configurations on ToneBench, side by side: Writing Elo, Writing Score, the nine metrics, and cost and time per script. A configuration is one model at one thinking level.

Highest Writing Elo
Elo 2704Claude Opus 5.5 (xhigh)#1 of 87 models · $1.00 per script
Lowest Cost
$0.0045 per scriptClaude Haiku 5.5 (low)Configuration #96 · Elo 1432

Recommended Configurations

Ranks in these notes are configuration ranks: each setting's place among all 151 configurations. The Highest Writing Elo card above gives the model rank.

Picks and Notes From Louis

For the strongest scripts, I'd use Claude Opus 5.5 (xhigh). It costs $1.00 and takes 6.7 minutes per script. I wouldn't pay for the max setting here: Claude Opus 5.5 (max effort) costs 2.6 times as much, takes 17.5 minutes, and trails xhigh by 5 Elo on the point estimate.

The new Claude Fable 5.2 is a big step up from Claude Fable 5.1, and at xhigh it gets closer to Opus than any Fable before it. It still doesn't change my pick. Claude Fable 5.2 (xhigh) trails Claude Opus 5.5 (xhigh) by 31 Elo on the point estimates and costs 2.2 times as much, for about the same wait. If you're on Claude Fable 5.2 anyway, xhigh is the setting I'd run.

The rest of that ladder is where people will overpay. Claude Fable 5.2 (max effort) lands behind xhigh at #4, and at $5.75 and 19.9 minutes a script, it's the most expensive and the slowest configuration on the whole board. The high setting, Claude Fable 5.2 (high), is #5 for $1.22, but Claude Opus 5.5 (xhigh) ranks above it, with ranges that don't overlap, for less money. And set the effort yourself: Anthropic says Claude Code now runs Claude Fable 5.2 at medium by default, and Claude Fable 5.2 (medium) drops to #11, behind Claude Opus 5.5 (high) on the point estimates, for more money.

Most days, though, I'd run Claude Opus 5.5 (high). It's #8 for $0.328 and comes back in 2.1 minutes, 233 Elo behind the top. For volume, my budget pick is Claude Sonnet 5.5 (high) at $0.092, still #35. And if you want to go really cheap, Claude Haiku 5.5 (xhigh) is #56 for $0.029 a script. Dropping all the way to Claude Haiku 5.5 (low) saves you pennies and lands you at #96.

Two I'd skip. Claude Sonnet 5.5 (max effort) costs more than Claude Opus 5.5 (xhigh) and ranks below it. And Claude Fable 5.1 (max effort) costs $3.56 and takes 12.3 minutes a script, while Claude Fable 5.2 (xhigh) ranks above it, with ranges that don't overlap, for less money and less waiting. If you're on the older Fable, I'd switch. Even Claude Fable 5.2 (high) ranks above it, with overlapping ranges, for $1.22 in 4.0 minutes, and both of those ranges also overlap that of Claude Opus 5.5 (high), which costs $0.328.

Claude Opus 5.5 (xhigh) is #1 by Writing Elo, and right behind it at #2 is Claude Opus 5.5 (max effort), the same model one setting higher, with a range that overlaps it. The new Claude Fable 5.2 (xhigh) comes next at #3, with a range that overlaps both of theirs, though overlap alone doesn't test their difference. Either way, the top of the board belongs to this family: you have to go down to #13 to find the first model from another lab. So if you write scripts with AI, this is the family I'd look at first. The hard part is picking one: 34 configurations of 13 models, spread over 1935 Elo, and the effort setting moves a model a long way. Claude Sonnet 5.5 (max effort) lands at #6, and with the effort turned down to low, the same model falls to #66. One disclosure: Claude Opus 5.5 is also one of the four judges, and the other three score every draft too.

All 34 Anthropic Configurations, Grouped by Model

Every row wrote the same 10 scripts for the same four blind judges, so these Writing Elo ratings sit on the leaderboard's scale. The rank column is the configuration rank, each setting's place among all 151 configurations on the board. Rows are grouped by model: a highlighted row is a model's best configuration, and the rows under it are the same model at other thinking levels. The ± after each Elo is half the width of its 95% range, and scores run from 0 to 100. Cost is one full script at the API list price of the route used. Each name opens its scorecard, and the head-to-head comparisons cover models from different labs.

Score colors:82 and up65 to 81.9under 65

With the metric scores shown, the table scrolls sideways.

Configuration RankConfigurationWriting EloWriting ScoreCost / ScriptTime / ScriptTone & Voice MatchWriting Craft & ClaritySubstance, Accuracy & ValueHook StrengthYouTube Best PracticesContinuity & EmotionSlop ScoreLength AdherenceVisual Cue Quality
#1Claude Opus 5.5 (xhigh)2704 ±5185.8$1.006.7 min82.583.288.482.984.485.488.598.684.9
#2Claude Opus 5.5 (max effort)2699 ±5885.9$2.5817.5 min82.882.988.083.784.385.788.498.785.9
#8Claude Opus 5.5 (high)2471 ±5084.6$0.3282.1 min81.582.687.382.783.385.088.092.783.1
#9Claude Opus 5.5 (medium)2452 ±5584.5$0.3101.9 min81.282.687.182.283.684.987.892.782.8
#10Claude Opus 5.5 (low)2425 ±6084.3$0.2431.5 min81.182.287.282.582.985.087.791.382.7
#3Claude Fable 5.2 (xhigh)2673 ±4385.6$2.227.4 min81.882.588.284.084.284.688.299.085.3
#4Claude Fable 5.2 (max effort)2646 ±4585.5$5.7519.9 min81.782.488.183.784.584.788.098.985.4
#5Claude Fable 5.2 (high)2511 ±5084.8$1.224.0 min81.082.187.582.383.384.188.098.684.0
#11Claude Fable 5.2 (medium)2422 ±5084.3$0.5611.8 min80.782.087.183.683.484.188.091.783.0
#12Claude Fable 5.2 (low)2415 ±4284.1$0.4601.4 min80.881.887.083.683.684.087.890.582.9
#6Claude Sonnet 5.5 (max effort)2496 ±5284.6$1.3814.3 min80.581.887.881.082.684.587.898.884.4
#15Claude Sonnet 5.5 (xhigh)2262 ±4283.2$0.3023.1 min79.580.886.681.281.882.786.594.580.2
#35Claude Sonnet 5.5 (high)1956 ±4481.0$0.0921.0 min77.579.084.581.880.581.084.686.075.3
#62Claude Sonnet 5.5 (medium)1776 ±4179.4$0.0650.6 min77.278.183.082.477.479.883.579.173.2
#66Claude Sonnet 5.5 (low)1730 ±3578.9$0.0610.7 min76.677.883.283.076.779.383.575.773.3
#7Claude Fable 5.1 (max effort)2485 ±3884.5$3.5612.3 min81.880.886.082.083.784.586.898.382.7
#14Claude Fable 5.1 (high)2282 ±6283.4$1.856.3 min80.879.885.180.982.783.785.695.880.9
#16Claude Opus 5 (max effort)2255 ±5483.2$0.7365.5 min80.880.584.582.383.583.785.192.079.8
#20Claude Opus 5 (high)2147 ±5282.6$0.4723.5 min80.380.284.381.882.083.184.690.577.8
#17Claude Fable 5 (max effort, Opus 4.8 fallback)2165 ±5882.9$3.2012.5 min80.677.584.978.084.384.484.694.781.2
#28Claude Fable 5 (high)2070 ±5181.9$1.344.9 min79.676.883.877.383.083.083.895.078.0
#39Claude Opus 4.8 (max effort)1926 ±3580.7$1.239.8 min76.776.782.080.380.480.982.194.679.7
#59Claude Opus 4.8 (high)1789 ±4079.3$0.1931.4 min75.977.081.381.876.480.280.788.574.3
#42Claude Haiku 5.5 (max effort)1914 ±4380.8$0.10312.6 min76.477.885.678.276.878.786.097.380.9
#56Claude Haiku 5.5 (xhigh)1846 ±4280.0$0.0293.6 min75.777.484.577.576.278.086.095.579.0
#71Claude Haiku 5.5 (high)1648 ±4678.0$0.0121.5 min74.777.282.378.372.877.485.484.972.9
#84Claude Haiku 5.5 (medium)1551 ±4476.7$0.00670.8 min74.076.781.579.171.176.484.877.671.6
#96Claude Haiku 5.5 (low)1432 ±4675.2$0.00450.5 min72.375.280.779.165.575.783.577.970.7
#61Claude Sonnet 5 (max effort)1779 ±4879.2$0.69612.0 min75.076.479.975.578.480.382.896.674.2
#83Claude Sonnet 5 (high)1561 ±5576.6$0.1211.9 min73.975.778.475.373.579.482.384.765.2
#69Claude Opus 4.7 (xhigh)1682 ±4378.3$0.3812.8 min75.975.779.279.275.378.081.390.673.6
#79Claude Opus 4.6 (high)1576 ±4676.9$0.7778.8 min73.573.078.873.975.376.780.693.573.9
#92Claude Sonnet 4.6 (high)1463 ±4676.0$0.66410.7 min71.573.778.275.373.875.280.390.672.8
#126Claude Haiku 4.5 (extended thinking)769 ±6665.4$0.0651.8 min62.264.666.373.362.166.672.867.851.8

Effort Ladders for Anthropic Models

Each of these models ran at more than one thinking level. Each page shows how the score, cost and time change from one level to the next.

Claude Opus 5.5 Against Other Anthropic Models

Each comparison lists every thinking level of both models, with Writing Elo, cost and time side by side. The line under each pair gives the Writing Elo gap between the two best configurations. A pair outside the written match-ups opens in the comparison tool.

← All comparisons