BENCH / TOWER DEFENSE IN PAUSED TIME

Ravelin benchmark

Ravelin, the daily tower defense, played with the game's own orders while the clock is paused. The bench runs the game's own simulation, so the boards are the ones people play: the daily sieges of fixed dates, campaign levels, and Iron levels the game's own bot cannot hold. A battle scores against par, the game's reference clear on the same board. Compare wins, lives lost, refused orders, and the cost and time of a decision.

16 model settings81 battles866 decisionsUpdated 2026-11-21

100 is winning every board at par, the game's own reference clear. A model above 100 won faster or lost fewer lives than par. Iron boards pull everyone down: the reference bot holds none of them. Paused-time scores compare with each other, not with a person playing the web game against the clock.

Ravelin leaderboard

Select a column to sort. Scroll sideways for all metrics →

#ModelScore95% intervalScore per tierDayCmpIrnWonBeat parLives lostRefusedDecisionsBattlesAPI cost / battleSec / decision
1
Gemini 4 Neonreasoning medium
107133131575 of 65 of 60.20.6%116$0.88740.4
2
Gemini 4 Hydrogenreasoning medium
99120119575 of 65 of 61.70.7%116$0.32716.6
3
GPT-6.2 Astrareasoning medium
89130124134 of 64 of 61.34.7%106$0.32313.6
4
GPT-6.1 Solreasoning medium
8713911674 of 64 of 60.54.3%96$0.13814.0
5
GPT-6.1 Terrareasoning medium
86136109134 of 64 of 61.01.4%106$0.0477.2
6
GPT-6 Solreasoning medium
8313510854 of 64 of 60.80.7%96$0.14016.0
7
GPT-6 Astrareasoning medium
8213011254 of 64 of 60.87.8%96$0.73412.2
8
Qwen3.8 Flash-Nextlocal, thinking low · open weights
731318534 of 63 of 64.30.8%116n/a587.5
9
Claude Haiku 5.5reasoning medium
7211110234 of 63 of 62.20.0%106$0.0139.9
10
Claude Sonnet 5.5reasoning medium
701099754 of 62 of 63.80.0%126$0.1225.0
11
GPT-6 Lunareasoning medium
57739533 of 62 of 66.015.4%116$0.00809.8
12
Gemini 4 Argonreasoning medium
843 of 612312562 of 32 of 33.01.4%103$0.1299.7
13
Qwen3.8 27Blocal, no thinking · open weights
383 of 6179261 of 30 of 310.037.0%123n/a148.3
14
Ministral 3 8Blocal, no thinking · open weights
253 of 6383160 of 30 of 313.762.1%163n/a40.7
15
Qwen3.5 9Blocal, no thinking · open weights
173 of 6212560 of 30 of 313.743.1%123n/a34.9
16
Gemma 3 4Blocal, no thinking · open weights
143 of 6171960 of 30 of 313.788.1%123n/a37.9

Score is 100 times the mean points over every battle a model played: a win scores par divided by its score (battle time plus 5 seconds per life lost), never below 0.5; a loss scores half the share of the waves it held. A model ranks once it has played the first 2 battles of every tier; until then it shows how many of those it has played and ranks after the models that have. Past that the bench keeps playing further battles in the same order for every model, and every finished one counts, so n grows and the intervals narrow.

Cost frontierAPI-equivalent cost per battle. Local models have no list price and are omitted. Bars show 95% intervals.
Closed weightsOpen weightsPareto frontier
050100150$0.010$0.030$0.100$0.300$1.00API-equivalent USD / battle (log scale)ScoreGemini 4 NeonGemini 4 HydrogenGPT-6.2 Astra
Time frontierObserved response time, including reasoning and queue wait. Routes and hardware differ.
Closed weightsOpen weightsPareto frontier
0501001500200400600Seconds / decisionScoreGemini 4 NeonGPT-6.2 Astra

The Pareto frontier shows where no other model scores at least as well for less cost or time, with one strict improvement. Partial sets are included and may cover different items. Use completed sets for direct comparisons.

By tier

Score against par with its 95% interval, then battles won, lives lost per battle, and how a won battle's score compared with par (below 1 is faster). The count after each result is n, the battles finished in that tier.

ModelDaily siegeCampaignIron
Gemini 4 Neonreasoning medium133 (99 to 166)n 2 · 2 won · 0.0 lives lost · 0.77x par when won131 (100 to 163)n 2 · 2 won · 0.0 lives lost · 0.77x par when won57 (0 to 156)n 2 · 1 won · 0.5 lives lost · 0.93x par when won
Gemini 4 Hydrogenreasoning medium120 (110 to 130)n 2 · 2 won · 4.0 lives lost · 0.83x par when won119 (116 to 121)n 2 · 2 won · 0.5 lives lost · 0.84x par when won57 (0 to 159)n 2 · 1 won · 0.5 lives lost · 0.92x par when won
GPT-6.2 Astrareasoning medium130 (103 to 158)n 2 · 2 won · 3.0 lives lost · 0.78x par when won124 (118 to 130)n 2 · 2 won · 0.0 lives lost · 0.81x par when won13 (0 to 31)n 2 · 0 won · 1.0 lives lost
GPT-6.1 Solreasoning medium139 (111 to 167)n 2 · 2 won · 0.0 lives lost · 0.73x par when won116 (105 to 128)n 2 · 2 won · 0.5 lives lost · 0.86x par when won7 (4 to 10)n 2 · 0 won · 1.0 lives lost
GPT-6.1 Terrareasoning medium136 (118 to 154)n 2 · 2 won · 2.0 lives lost · 0.74x par when won109 (95 to 123)n 2 · 2 won · 0.0 lives lost · 0.92x par when won13 (0 to 31)n 2 · 0 won · 1.0 lives lost
GPT-6 Solreasoning medium135 (94 to 176)n 2 · 2 won · 0.0 lives lost · 0.76x par when won108 (101 to 116)n 2 · 2 won · 1.5 lives lost · 0.92x par when won5 (4 to 6)n 2 · 0 won · 1.0 lives lost
GPT-6 Astrareasoning medium130 (95 to 165)n 2 · 2 won · 1.0 lives lost · 0.79x par when won112 (108 to 115)n 2 · 2 won · 0.5 lives lost · 0.90x par when won5 (4 to 6)n 2 · 0 won · 1.0 lives lost
Qwen3.8 Flash-Nextlocal, thinking low131 (96 to 166)n 2 · 2 won · 1.0 lives lost · 0.78x par when won85 (55 to 115)n 2 · 2 won · 11.0 lives lost · 1.21x par when won3 (0 to 8)n 2 · 0 won · 1.0 lives lost
Claude Haiku 5.5reasoning medium111 (92 to 130)n 2 · 2 won · 4.5 lives lost · 0.91x par when won102 (86 to 117)n 2 · 2 won · 1.0 lives lost · 0.99x par when won3 (0 to 8)n 2 · 0 won · 1.0 lives lost
Claude Sonnet 5.5reasoning medium109 (93 to 125)n 2 · 2 won · 5.5 lives lost · 0.92x par when won97 (92 to 102)n 2 · 2 won · 5.0 lives lost · 1.03x par when won5 (4 to 6)n 2 · 0 won · 1.0 lives lost
GPT-6 Lunareasoning medium73 (0 to 175)n 2 · 1 won · 11.0 lives lost · 0.80x par when won95 (82 to 108)n 2 · 2 won · 6.0 lives lost · 1.06x par when won3 (0 to 8)n 2 · 0 won · 1.0 lives lost
Gemini 4 Argonreasoning medium123n 1 · 1 won · 8.0 lives lost · 0.82x par when won125n 1 · 1 won · 0.0 lives lost · 0.80x par when won6n 1 · 0 won · 1.0 lives lost
Qwen3.8 27Blocal, no thinking17n 1 · 0 won · 20.0 lives lost92n 1 · 1 won · 9.0 lives lost · 1.08x par when won6n 1 · 0 won · 1.0 lives lost
Ministral 3 8Blocal, no thinking38n 1 · 0 won · 20.0 lives lost31n 1 · 0 won · 20.0 lives lost6n 1 · 0 won · 1.0 lives lost
Qwen3.5 9Blocal, no thinking21n 1 · 0 won · 20.0 lives lost25n 1 · 0 won · 20.0 lives lost6n 1 · 0 won · 1.0 lives lost
Gemma 3 4Blocal, no thinking17n 1 · 0 won · 20.0 lives lost19n 1 · 0 won · 20.0 lives lost6n 1 · 0 won · 1.0 lives lost

The game's bot on the same boards

PlayerScoreDaily siegeCampaignIron
Par (reference bot)The game's own reference bot at its own pace. Its score on a board is the par a player sees.764 of 6 won1002 won1002 won280 won
Rusher botThe same bot calling every wave the moment it may, for the fastest clears.1144 of 6 won1482 won1652 won300 won
Bot in paused timeThe same bot held to the bench's decision points, as many orders as it likes at each.583 of 6 won641 won902 won190 won
Novice botThe same bot playing plainly: no abilities, never calls a wave early.634 of 6 won972 won852 won70 won
Every board

Every battle

One square per board, filled when the battle held every wave. Columns run tier by tier, in the order every model meets them.

MethodologyProtocol, scoring & limitations

The boards

Ravelin is a tower defense on a war table: attackers march along the road from the gates, and towers built beside it shoot them before they reach the keep. Each attacker that gets through costs lives. The bench runs Ravelin's own simulation, pinned to commit 23c8ab1, so a later balance change in the game cannot move anyone's result. Every battle stores the game's own tick tape, and the same board and orders reproduce the same battle, in the bench's driver and in the game's own replayer.

TierBoardsItemsReference botRusher bot
Daily siegeThe daily board of a fixed date, as every player got it6100148
CampaignA campaign level, chapters I to III6100165
IronAn Iron level: one life, 35% tougher waves, no abilities32830

Every model gets the same boards in the same order from set v1: the daily sieges of six fixed dates, six campaign levels across the game's three chapters, and three Iron levels (one life, waves 35% tougher, no abilities). A model ranks once it has played the first 2 of every tier. After that the bench keeps playing further boards, up to 6 per tier.

A decision

Play is in paused time. Every decision is a fresh request with the whole board as text: the map with the towers drawn on it, gold, lives, every tower with its stats and upgrade costs, the attackers on the field, the next three waves (the same full intel the game shows), and what happened since the last decision, including any refused orders and why. The model answers with up to 12 orders: build, upgrade (choosing a final form at tier 3), sell, retarget, the two abilities, and calling the next wave early. The first decision is planning, with the clock stopped. After that the battle runs and stops for the next decision when the field is clear, when a life is lost, or after 20 seconds of battle. After 150 decisions the battle runs to its end without further orders.

An order the game refuses is not retried; it is counted and reported back, as the game tells a player. The one retry is for a reply with no readable orders tag. A second unreadable reply waits.

Score

Ravelin scores a won battle by its time plus 5 seconds per life lost, lower is better, and shows a par: its reference bot's score on that board. Here a won battle scores par divided by its score, never below 0.5, so beating par scores above 1. A lost battle scores half the share of the waves it held, so any win beats any loss. An Iron board takes its normal version's par, since the bot cannot hold the Iron one. The benchmark score is 100 times the mean, so 100 is winning every board at par. The 95% interval is the ratio estimator's, by the delta method. The game's bot is shown on the same boards in four settings for scale.

Cost and time

Same as the chess and sudoku benchmarks: Claude models run through Claude Code, OpenAI models through the Codex CLI and Google models through the Antigravity CLI on a subscription, with every tool switched off, and local models on our own DGX Sparks. Cost is API-equivalent: what the tokens of every request in the battle, retries included, would cost at the lab's list price. Local models have no list price, so they show n/a, not zero. Time is the wall-clock wait for the replies, including any queue on our side, so compare it within a route rather than across them. A request that fails on our side (a timeout, an outage, a usage limit) aborts the game, and an aborted game never counts against the model.

What it doesn't show

Paused time removes reaction speed on purpose: the model is asked again at every clear, every lost life and every 20 seconds, and never has to act mid-thought. These scores compare with each other and with the bot, not with a person playing the web game against the clock.

Data: JSON · CSVUse with credit to Glimmer Index.