Towards AITowards AIToneBench

Model Comparison

Kimi K3 vs Claude Haiku 5.5

Kimi K3 (max) leads Claude Haiku 5.5 (max effort) by 202 Elo, and their 95% ranges sit 116 Elo apart. Among 87 models, each ranked by its best configuration, Kimi K3 is #12 and Claude Haiku 5.5 is #20. Cost per script: $0.362 for Kimi K3 (max) and $0.103 for Claude Haiku 5.5 (max effort).

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

Kimi K3
2116 Writing Elo of Kimi K3 (max), its best configuration#12 of 87 models · 82.2/100 Writing Score · $0.362 per script2 configurations on the board
Claude Haiku 5.5
1914 Writing Elo of Claude Haiku 5.5 (max effort), its best configuration#20 of 87 models · 80.8/100 Writing Score · $0.103 per script5 configurations on the board

Picks and Verdict

PickKimi K3

Kimi K3 (max)

Pick Kimi K3 (max) for scripts you plan to film, since it leads Haiku's best by 202 Elo with separate ranges for $0.362 a script.

PickClaude Haiku 5.5

Claude Haiku 5.5 (xhigh)

Pick Claude Haiku 5.5 (xhigh) for cheap drafts at $0.029, since its range overlaps Kimi K3 (high) at about 4.4 times the price.

At the top, Kimi K3 (max) is the clear pick: 202 Elo ahead of Claude Haiku 5.5 (max effort), with ranges that don't touch. Ranked by best setting, that's the #12 model of 87 against the #20 model. On a read-through, I'd notice the hooks first, about five points higher on Kimi, then YouTube structure, more than four. Haiku has real strengths of its own: better substance, length adherence and slop scores, and slightly steadier drafts. One level down, it gets closer. Kimi K3 (high) sits 23 Elo under Haiku's max effort and 45 above Haiku at xhigh, and its range overlaps both. Haiku at xhigh costs $0.029 a script against $0.128 for Kimi at high, less than a quarter of the price. Haiku's max effort is also cheaper than Kimi at high, but it takes 12.6 minutes, the slowest setting on this page. Under xhigh, Haiku falls off fast, so I wouldn't go lower. For a script I'm filming, I'd pay $0.362 for Kimi at max. For bulk drafting, Haiku at xhigh is the better deal.

Disclosure: Claude Haiku 5.5 comes from the same lab as one of the four judges, Claude Opus 5.5. The other three sit outside its lab, so it is never scored by its own lab alone (how the panel works). The judge analysis measures how much each judge favors its own lab.

How to Read the Writing Score, Writing Elo and 95% Ranges

Writing Score and Writing Elo. The Writing Score (0 to 100) is the weighted average of the nine metric scores: seven from the judges, Length Adherence from code and the Slop Score half from each. The Writing Elo is a rating from pairwise comparisons: on each script, each configuration's average across its drafts is compared with every other configuration's, and gaps within the draft-to-draft variation count as draws. Elo rewards winning often, so the two can put configurations in a different order.

Elo ranges. Each 95% range shows how far a configuration's Elo could move if the drafts had come out slightly differently. It comes from resampling every configuration's recorded draft scores on each reference script and ranking again. Comparing two separate ranges does not test the Elo gap between them: ranges far apart are strong evidence for the order, and overlapping ranges do not mean a tie. The reference scripts stay fixed, so the ranges say nothing about new scripts. The methodology page has the full description.

All Thinking Levels of Both Models

Every ranked configuration of Kimi K3 and Claude Haiku 5.5, scored on the same 10 scripts by the same four judges.

How to Read the Chart
  • Each row is one configuration: the model at one thinking level. The dot is its Writing Elo, a rating from pairwise comparisons of the scores on each script, and the bar is its 95% range: how far that Elo could move if the drafts had come out slightly differently.
  • Each bar describes one configuration. Overlap alone does not establish a tie or decide their order. Comparing these ranges does not directly test the difference between a pair.
  • The table under the chart lists the same configurations in the same order and adds the Writing Score (the 0 to 100 consensus of the judges and code), cost, time and configuration rank (the place among all configurations on the board).
Kimi K3Claude Haiku 5.5
ConfigurationWriting Elo95% RangeWriting ScoreCost / ScriptTime / ScriptReasoning TokensConfiguration Rank
Kimi K3 (max)21162076 to 215282.2$0.3629.9 min18,568#26
Claude Haiku 5.5 (max effort)19141874 to 196080.8$0.10312.6 minN/A#42
Kimi K3 (high)18911838 to 193880.3$0.1282.8 min3,081#45
Claude Haiku 5.5 (xhigh)18461804 to 188780.0$0.0293.6 minN/A#56
Claude Haiku 5.5 (high)16481603 to 169578.0$0.0121.5 minN/A#71
Claude Haiku 5.5 (medium)15511505 to 159376.7$0.00670.8 minN/A#84
Claude Haiku 5.5 (low)14321383 to 147575.2$0.00450.5 minN/A#96

Cost is one script at the API list price of the route used, uncached and without judging. Time is the average for one accepted script, retries included. N/A under Reasoning Tokens means the board has no separate reasoning count for that route.

Computed Findings Across Thinking Levels

Each finding is computed from the configurations above. The 95% ranges describe individual configurations; overlap alone establishes neither a tie nor an order for a pair.

Writing Elo at Each Shared Level Name

LevelKimi K3Claude Haiku 5.5Gap95% Ranges
high18911648+243Separate
max21161914+202Separate

Gap is Kimi K3 minus Claude Haiku 5.5 at the same level name, using each model's highest-rated configuration at that level.

Two Chosen Configurations Compared

Pick one configuration of each model to compare their metric scores, consistency across drafts and per-script scores. Each menu starts on the model's best configuration.

Scores by Metric

Kimi K3 (max)Claude Haiku 5.5 (max effort)Solid bar: higher score in that row

Kimi K3 (max) scores higher on five of the nine metrics and Claude Haiku 5.5 (max effort) on four. Weighted by each metric's share of the Writing Score, YouTube Best Practices does the most to put Kimi K3 (max) ahead, with a 4.4-point lead on that metric. The weight beside each metric is its share of the Writing Score.

Tone & Voice Match19% weight
78.7
76.4
Writing Craft & Clarity13% weight
78.8
77.8
Substance, Accuracy & Value15% weight
83.5
85.6
Continuity & Emotion14% weight
82.3
78.7
YouTube Best Practices12% weight
81.2
76.8
Hook Strength10% weight
83.1
78.2
Length Adherence8% weight
94.2
97.3
Slop Score5% weight
83.4
86.0
Visual Cue Quality4% weight
80.0
80.9

Consistency Across Drafts

Kimi K3 (max)Claude Haiku 5.5 (max effort)
Score spread across drafts (±)1.681.55
Lowest script score80.0 Script 1078.2 Script 10
Highest script score83.8 Script 982.6 Script 4
Range across scripts3.84.4

Spread is how much the Writing Score moves between a configuration's drafts, and lower is steadier. Green marks the better number in each row.

Scores by Script

Writing Score on each reference script. Gap is Kimi K3 (max) minus Claude Haiku 5.5 (max effort), and the higher score is in bold.

ScriptKimi K3 (max)Claude Haiku 5.5 (max effort)Gap
Script 1 opinion / warning explainer81.480.3+1.1
Script 2 news-analysis / skeptical explainer82.680.7+1.9
Script 3 personal roadmap / opinion81.880.8+1.0
Script 4 short explainer80.182.6-2.5
Script 5 news-analysis / opinion explainer83.481.7+1.7
Script 6 founder announcement / personal origin story83.281.1+2.1
Script 7 personal technical walkthrough / agentic workflow case study83.680.0+3.6
Script 8 product release announcement / personal observation82.381.3+1.0
Script 9 career guide / hiring analysis83.881.5+2.3
Script 10 engineering process walkthrough / presentation adaptation80.078.2+1.8

Compare With Another Model

Pick a second model. A pair with its own written match-up opens that page, and any other pair opens in the comparison tool.

Nearby Match-Ups by Best Writing Elo

Every configuration wrote five drafts for each of the same 10 briefs, from the same prompt. Generation was closed-book, with the prompt as the only source and no web search or other tools. Each judge scores every draft once: Claude Opus 5.5 at medium effort through the Claude Code subscription, GPT-6.1 Sol at medium reasoning through the Codex CLI, GLM 5.3 Flash through OpenRouter, and Jev 1.13, a typed judge from TypeSafe that answers each metric as a multiple-choice question. The channel's finished script is their reference for the voice, and it gets no score on the board itself.

← All comparisons