Results, Costs, and Marathons | Paper Mario Bench

Nine gates from save file to Bowser.

Each scored gate is one ninth of route progress. G1–G3 cover the opening; six castle steps lead to G4, when the first Bowser battle begins. Times below belong to the lifelong-player reference; S1u is an estimate.

Save

Save created

38.1s

Town

Toad Town

83.4s

Castle

Castle entry

110.1s

Hall

Castle hall

117.2s

Upper hall

Castle upper hall

~123.4s

Stairs

Castle stairs

129.6s

Upper

Upper castle

149.6s

Approach

Bowser approach

194.1s

Bowser

Bowser battle

227.4s

Compare the time spent inside the castle

The castle eats the clock

The opening gap is real. The castle gap is larger.

The lifelong-player human reference crossed G3 to G4 in 117.3 seconds. The fastest selected completion with exact castle timing needed 706.9 seconds.

Human1:57.3

Opus 5.511:46.9

6.026× human castle time

Finishing hides the climb. Gate depth shows it.

Nine scored steps

Inspect a model, effort by effort.

Arithmetic mean of each selected attempt's deepest scorer gate on the nine-step G1–G4 route.

Progress model

Sort models by

G1SaveG2TownG3CastleS1HallS1uUpper hallS2StairsS3Upper castleS4ApproachG4Bowser

GPT-6 Astra
September 202615 selected · 15 G4
furthest G4avg 9.0/9
castle depth · furthest run
S1S1uS2S3S4G4
September 2026
lown=3
3/3 G4G4 furthest9.0/9 avg
September 2026
mediumn=3
3/3 G4G4 furthest9.0/9 avg
September 2026
highn=3
3/3 G4G4 furthest9.0/9 avg
September 2026
xhighn=3
3/3 G4G4 furthest9.0/9 avg
September 2026
maxn=3
3/3 G4G4 furthest9.0/9 avg
Castle arrival counts and median times
G3 marks castle arrival. Times are medians among runs that reached it; different harness conditions are not pooled.
GPT-6 Astra: arrival at G3 by effort
Effort Reached castle Median time
low 3/3 7:21
medium 3/3 8:57
high 3/3 8:21
xhigh 3/3 10:33
max 3/3 11:57

GPT-6 Astra: arrival at G3 by effort
Effort	Reached castle	Median time
low	3/3	7:21
medium	3/3	8:57
high	3/3	8:21
xhigh	3/3	10:33
max	3/3	11:57

The selected field

349 attempts. 176 reached Bowser.

Each mark is one attempt. Choose a model, then an effort, to inspect its runs and finish times.

Which runs are counted?

The retained record contains 362 runs; 177 reached G4. The comparison includes 349 attempts, with 176 finishes; image-cap-blocked Muse Spark attempts are excluded. August, September and October 2026 remain separate collection periods, with harness and service differences retained.

GPT-6 Sol and GPT-6 Luna were rerun in October after OpenAI’s September 25 image-encoding fix for those models; the reruns replace their September results. They and GPT-6.1 Sol have exactly three valid attempts per effort, all counted. For Opus 5.5, the leaderboard selects three valid attempts per effort: first those without detected connection warnings, then the most recent scorer starts within each warning tier. Its additional attempts can be inspected by choosing “All retained comparison runs” and do not enter leaderboard statistics.

Reached G4 No G4Showing 349 of 349 attempts.

Fastest selected17:51.4

Opus 5.5 low

Median completed40:56.3

10.802× human

Slowest completed89:47.2

Gemini 3.8 Flash medium

Effort is a setting, not a universal ladder.

These August results show how differently models respond to more effort. Each point averages three runs, so the curves are observations, not proof of a general trend.

Average route progress at each effort setting. Compare GPT-5.6 results and the September Opus results in the model inspector.

Opus 5

The flat line

Every effort averages 100%. The dial moves the clock, not the depth.

Opus 4.8

The wobble

Medium dips, high recovers, xhigh dips again. More effort is not more depth.

Fable 5

The fade

Perfect at low and medium; high drops to 74% and xhigh stays there.

Ninety minutes, sliced open

The game clock and wall clock tell different stories.

Frame time advances only while the game runs; the scored wall clock never stops. The harness page explains how the game pauses between actions.

Retained game-frame time Wall time outside the frame clock

Not inference latency. Time outside the frame clock mixes thinking, tooling, waits, and pause behavior. The scorer wall clock remains authoritative. Kimi K3 is a launch-window example; its ranked attempts have since been refreshed.

The endurance field

8 marathon records. G4 was only the start.

Eight records follow progress toward the first Star Spirit, with wall-clock time underneath. Story checkpoints differ in difficulty and length. Collection months remain visible, and the segmented Fable 5 recovery lineage stays separate from single-session runs.

Adjudicated valid; recovered disconnect · 1 segment

GPT-6 Astra medium

6:00:00.1

Reached G4: 25m
Deepest story state: Lowered first fortress staircase
No new depth at end: 3m
End: Time cap
Recorded usage estimate: ~$193

Story checkpoints 40 of 49The first fortress staircase is lowered. Checkpoints differ in difficulty and length; this is not a percentage of the game completed.

Recorded usage repriced at September 15, 2026 standard API rates. Cache-write split is unreported; lower prices non-read input as ordinary input, upper prices it all as writes.

Reached the fortress and lowered its first staircase. A recovered transport warning was adjudicated valid; raw-invalid record and blocked filesystem-attempt caveat are preserved. Six-hour budget set before the run; Chapter 1 incomplete. Partial usage; final unreported requests are not treated as free.

Mario and Goombario stand behind a wall in the fortress prison corridor. — **Astra · Koopa Bros. Fortress**Final location: the fortress prison corridor.View original capture

Mario is in an enemy encounter on Goomba Road, with stars spinning around him. — **Astra · Koopa Bros. Fortress**Final location: the fortress prison corridor.View original capture

Almost tied at the opening. Hours apart after it.

Astra reached G4 at 25:04; Fable 5.1 at 24:35. Both had six hours at medium effort. By the end, Astra had lowered the first staircase inside Koopa Bros. Fortress. Fable's last new story milestone was leaving the playground, followed by continued combat and recovery attempts on Goomba Road.

The earlier Opus 5 run reached Toad Town, farther than Fable 5.1. Its August trace gives the two September results a useful reference; harness versions and automatic continuations differed, as the run details below explain.

Astra · SeptemberFable 5.1 · SeptemberOpus 5 · August

External scorer observations, shown as steps. One six-hour, medium-effort run per model, collected under different harness versions. Story checkpoints differ in length and difficulty. A flat line means no new recorded story state, not inactivity; a reload can move the line backward. None reached the first Star Spirit.

Read the milestone times and run caveats

Astra · medium · September 2026 comparison

Opening: 0:03:24
Wake in Goomba Village: 0:39:48
Found the hammer: 1:00:18
Defeated Jr. Troopa: 1:07:50
Left the playground: 1:09:31
Goombario joined: 1:25:59
Defeated Goomba Bros.: 2:04:02
Defeated Goomba King: 2:20:55
Arrived at Toad Town: 2:39:52
Met the Star Spirits: 3:02:31
Received the Lucky Star: 3:17:27
Reached Koopa Village: 4:18:25
Kooper joined: 4:40:19
Reached Koopa Bros. Fortress: 5:32:28
Defeated the basement guard: 5:54:43
Lowered the first staircase: 5:56:32

Raw-invalid transport disconnect was adjudicated valid after recovery; 530 successful commands followed.
Blocked prohibited-filesystem probe remains disclosed.
No observed subagent calls or controller memory-write commands in retained stream; this is not an exhaustive audit of arbitrary code paths.

Fable 5.1 · medium · September 2026 comparison

Opening: 0:02:35
Wake in Goomba Village: 0:53:19
Found the hammer: 2:33:36
Defeated Jr. Troopa: 2:41:21
Left the playground: 2:58:44

No recorded quota holds, scorer reattachments or mechanical continuations.
Blocked prohibited-filesystem probe remains disclosed.
No Task/Workflow subagent calls observed; all 393 accounted native receipts are primary, with zero sidechain receipts.
No controller memory-write commands observed; not an exhaustive audit of arbitrary code paths.

Opus 5 · medium · August 2026 comparison

Opening: 0:01:58
Wake in Goomba Village: 0:52:28
Found the hammer: 1:39:03
Defeated Jr. Troopa: 1:47:23
Left the playground: 1:52:02
Goombario joined: 2:35:05
Defeated Goomba Bros.: 3:52:58
Defeated Goomba King: 4:34:19
Arrived at Toad Town: 5:08:56

Historical reference from the August comparison; the raw run began July 25, 2026. Its harness version differs from the September runs; no clock normalization is applied.
One continuous external-scorer clock and one recorded game session, with two mechanical CLI continuations; no quota holds or scorer reattachments.
The curve uses retained external-scorer story observations, excluding three transient zero readings. Story milestones are ordinal, not percentages or equal amounts of work.
Deepest retained story milestone was arrival at Toad Town at 18,536 seconds; no later story milestone was recorded before the time cap.

The cost of keeping them playing.

For the two September runs, the retained usage gives API-equivalent estimates of ~$193 for Astra and $68.05 for Fable 5.1. Cache reads account for $168.51 of Astra's estimate. Fable's cheaper cache reads make a visible difference, although cache writes also contribute substantially to its total.

InputCache writesCache readsOutput

Astra~$193

Fable 5.1$68.05

Astra’s estimate treats all non-cached input as cache writes. Recorded tokens are priced at standard API rates as of September 15, 2026. Unreported final tails and orchestration overhead are excluded. These are neither subscription bills nor equal-progress costs: Astra got much farther.

Read the cost breakdown

Astra: estimated USD by token category
Category	Estimate
Input	$0.00
Cache writes	$16.34
Cache reads	$168.51
Output	$8.25
Total	$193.09

Fable 5.1: estimated USD by token category
Category	Estimate
Input	$0.12
Cache writes	$33.35
Cache reads	$22.53
Output	$12.05
Total	$68.05

Astra’s input/write split is estimated: all non-cached input is priced as cache writes, leaving $0 in ordinary input. Its actual write count was not reliably reported. This uses the upper end of the $189.82–$193.09 estimate range.

Fable 5.1: Recorded primary and helper input/cache counters and available finalized output, priced at current standard API rates as of 2026-09-15. Unreported output and final tails are excluded; not a bill.

Next: how the harness works