ROUTE 101
◀ Route 101 PILOT  SEP 28, 2026 · 1 SEED

Pilot

1 RUN · SEED 7 · EXPLORATORYOne run shows what happened on this save. It isn't a reliable ranking of these models yet; repeat seeds come next.

In one race on seed 7, Sonnet 5 and Gemini 3 Flash all reached Exit 4 (got a starter). Sonnet 5 got there first, on move 76. DeepSeek V4 Flash stopped at Exit 2 (set the clock). GPT-5 mini stopped at Exit 1 (left the truck).

The first test: four models, 120 moves each, before the harness was finished. Kept for the record.

Pilot final standings: Sonnet 5 exit 4, Gemini 3 Flash exit 4, DeepSeek V4 Flash exit 2, GPT-5 mini exit 1 ▶ VIDEO ON X
Race video · pinned on @harrisonjohnson
▶ Subscribe on YouTube @harrisonjohnson on X
Result · 1 run · seed 7 · exploratory

How far each model got

The furthest story exit wins. Ties go to whoever got there in fewer moves. Cost is tracked on its own and doesn't affect the rank.

#ModelProgressReached atCostCost per exit
1◆ Sonnet 5
Anthropic
OpenRouter default routing · host not logged		 ████░░░░░░░░
Exit 04 · Got a starter		 move 076$1.01 $0.25
2▲ Gemini 3 Flash
Google
OpenRouter default routing · host not logged		 ████░░░░░░░░
Exit 04 · Got a starter		 move 099$0.16 $0.041
3● DeepSeek V4 Flash
DeepSeek
OpenRouter default routing · host not logged		 ██░░░░░░░░░░
Exit 02 · Set the clock		 move 092$0.021 $0.011LOWEST
4■ GPT-5 mini
OpenAI
OpenRouter default routing · host not logged		 █░░░░░░░░░░░
Exit 01 · Left the truck		 move 004$0.17 $0.17
  • 1
    ◆ Sonnet 5 Anthropic
    OpenRouter default routing · host not logged
    ████░░░░░░░░
    Furthest
    Exit 4 · Got a starter
    Reached at
    move 76
    Cost
    $1.01
    Cost per exit
    $0.25
  • 2
    ▲ Gemini 3 Flash Google
    OpenRouter default routing · host not logged
    ████░░░░░░░░
    Furthest
    Exit 4 · Got a starter
    Reached at
    move 99
    Cost
    $0.16
    Cost per exit
    $0.041
  • 3
    ● DeepSeek V4 Flash DeepSeek
    OpenRouter default routing · host not logged
    ██░░░░░░░░░░
    Furthest
    Exit 2 · Set the clock
    Reached at
    move 92
    Cost
    $0.021
    Cost per exit
    $0.011 · lowest
  • 4
    ■ GPT-5 mini OpenAI
    OpenRouter default routing · host not logged
    █░░░░░░░░░░░
    Furthest
    Exit 1 · Left the truck
    Reached at
    move 4
    Cost
    $0.17
    Cost per exit
    $0.17

    • Cost per exit = total cost ÷ exits reached. It measures efficiency; the rank comes from exits. Pilot run on one seed, before the current harness.

Play stats

How each model played

Counted from each model's own log.

◆ Sonnet 5
Words of reasoning
1,842 (18.2 a move)
Thinking time
4.4s a move (7 min total)
Tokens a move
4,336 in · 110 out
Cost a move
$0.01
Battle moves
2
Moves without walking
43 of 102
Moves without acting
0
Notes rewritten
5
▲ Gemini 3 Flash
Words of reasoning
3,181 (26.5 a move)
Thinking time
2.1s a move (4 min total)
Tokens a move
2,321 in · 58 out
Cost a move
$0.001
Battle moves
2
Moves without walking
45 of 120
Moves without acting
0
Notes rewritten
6
● DeepSeek V4 Flash
Words of reasoning
2,470 (20.6 a move)
Thinking time
3.5s a move (7 min total)
Tokens a move
2,336 in · 157 out
Cost a move
$0.00
Battle moves
0
Moves without walking
25 of 120
Moves without acting
4
Notes rewritten
8
■ GPT-5 mini
Words of reasoning
3,345 (27.9 a move)
Thinking time
7.3s a move (15 min total)
Tokens a move
1,670 in · 478 out
Cost a move
$0.001
Battle moves
0
Moves without walking
16 of 120
Moves without acting
16
Notes rewritten
2
$0$0.25$0.50$0.75$1000020040060080100120◆ Sonnet 5 $0.99▲ Gemini 3 Flash $0.16● DeepSeek V4 Flash $0.021■ GPT-5 mini $0.16

Spend over the race, from the cost of each move. The totals in the result table are a little higher because they also include the memory the model rewrites every 50 moves (Sonnet 5 $0.022, Gemini 3 Flash $0.003, DeepSeek V4 Flash $0.00, GPT-5 mini $0.003). Moves without walking include reading dialogue and battles.

Route

Where each model stopped

Pilot · where each model stopped
◆ Sonnet 5▲ Gemini 3 Flash● DeepSeek V4 Flash■ GPT-5 mini
01Left the truck02Set the clock03Rescued Birch04Got a starter05Oldale Town06Beat the rival07Got the Pokédex08Petalburg City09Rustboro City10Stone Badge11Dewford Town12Knuckle BadgeGPT-5 mini: exit 1, move 4■DeepSeek V4 Flash: exit 2, move 92●Sonnet 5: exit 4, move 76◆Gemini 3 Flash: exit 4, move 99▲
  1. 01Left the truck■
  2. 02Set the clock●
  3. 03Rescued Birch
  4. 04Got a starter◆ ▲
  5. 05Oldale Town
  6. 06Beat the rival
  7. 07Got the Pokédex
  8. 08Petalburg City
  9. 09Rustboro City
  10. 10Stone Badge
  11. 11Dewford Town
  12. 12Knuckle Badge
Timeline

When each model hit each exit

Each marker is an exit. The line ends where the model's moves ran out.

000020040060080100120MOVE ▶◆ Sonnet 5Sonnet 5: exit 1 (Left the truck) at move 401Sonnet 5: exit 2 (Set the clock) at move 1502Sonnet 5: exit 3 (Rescued Birch) at move 7603Sonnet 5: exit 4 (Got a starter) at move 7604▲ Gemini 3 FlashGemini 3 Flash: exit 1 (Left the truck) at move 401Gemini 3 Flash: exit 2 (Set the clock) at move 2002Gemini 3 Flash: exit 3 (Rescued Birch) at move 9903Gemini 3 Flash: exit 4 (Got a starter) at move 9904● DeepSeek V4 FlashDeepSeek V4 Flash: exit 1 (Left the truck) at move 1601DeepSeek V4 Flash: exit 2 (Set the clock) at move 9202■ GPT-5 miniGPT-5 mini: exit 1 (Left the truck) at move 401
Moments

What they were thinking

Straight from each model's move-by-move notes.

◆ SONNET 5 The bag is open showing the Poke Ball with Torchic. I need to select it to release Torchic and save Prof. Birch.

First to Exit 4: Got a starter · move 76.

Method

How this race was run

Game
Pokémon Emerald, new game, seeded RNG, trainer ID and clock
Seed
7
Tools
pathfinding tools (walk to a tile, talk, read, pick from menus)
Move budget
120 per model
Safety cap
$1 per run (stops runaway loops; moves decide the race)
Model settings
each provider's default settings
Score
furthest story exit from the game's own flags, then fewest moves to reach it
Race ends
after 120 moves, or at the safety cap, or if the model's API fails 3 moves in a row
Hosting
OpenRouter default provider routing
Cost
cost billed by OpenRouter
Harness
These runs predate per-run version stamps. The code that plays the race hasn't changed since.
Runs
pilot.s7.claude-sonnet-5 pilot.s7.deepseek-v4-flash pilot.s7.gemini-3-flash-preview pilot.s7.gpt-5-mini

Moves where a model didn't act: DeepSeek V4 Flash 4 (4 cut off by a reply-length limit) · GPT-5 mini 16.

Evidence

Every move, every model

Each file has the model's reasoning, action and result for every move. Text the game showed on screen is replaced with [game text].

Share

Post this race