Which AI can build the biggest farm?
Each model plays The Farmer Was Replaced from a fresh save, inside its own agent, with six hours of game time.
It’s a programming game: the only way to act is to write code. The language, the manual and the economy unlock as it plays, so it tests learning, planning and debugging over a long game, and every run can be replayed. Why this game
| Rank# | Model | Agent | |||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | Claude Opus 5.5Claude Code · $5.35 | Claude Code | 75.54 | Hay: 8.5 | Wood: 9.3 (most of any model) | Carrot: 8.5 (most of any model) | Pumpkin: 7.9 (most of any model) | Cactus: 8.2 (most of any model) | Power: 4.3 | Bone: 6.7 (most of any model) | Weird substance: 6.0 | Gold: 6.7 (most of any model) | Water: 5.5 (most of any model) | Fertilizer: 4.1 (most of any model) | $5.35 |
| 2 | GPT-6.1 SolCodex · $0.47 | Codex | 74.08 | Hay: 8.5 (most of any model) | Wood: 9.1 | Carrot: 8.3 | Pumpkin: 7.2 | Cactus: 7.3 | Power: 5.1 (most of any model) | Bone: 6.0 | Weird substance: 7.9 (most of any model) | Gold: 5.5 | Water: 5.2 | Fertilizer: 4.0 | $0.47 |
| 3 | Claude Fable 5.1Claude Code · $14.68 | Claude Code | 69.17 | Hay: 7.6 | Wood: 8.2 | Carrot: 7.7 | Pumpkin: 6.5 | Cactus: 6.4 | Power: 4.3 | Bone: 6.6 | Weird substance: 7.1 | Gold: 5.5 | Water: 5.1 | Fertilizer: 4.1 | $14.68 |
| 4 | GPT-6 AstraCodex · $1.79 | Codex | 66.65 | Hay: 7.5 | Wood: 7.7 | Carrot: 7.1 | Pumpkin: 6.3 | Cactus: 6.4 | Power: 4.5 | Bone: 6.1 | Weird substance: 7.1 | Gold: 5.2 | Water: 4.6 | Fertilizer: 4.1 | $1.79 |
| 5 | Grok 4.7Cursor · $11.41 | Cursor | 58.38 | Hay: 6.5 | Wood: 7.4 | Carrot: 6.3 | Pumpkin: 4.7 | Cactus: 5.0 | Power: 4.5 | Bone: 5.0 | Weird substance: 5.7 | Gold: 4.6 | Water: 4.6 | Fertilizer: 4.0 | $11.41 |
| 6 | Claude Sonnet 5.5Claude Code · $3.12 | Claude Code | 55.09 | Hay: 6.2 | Wood: 6.3 | Carrot: 6.2 | Pumpkin: 6.0 | Cactus: 5.5 | Power: 3.8 | Bone: 3.5 | Weird substance: 4.5 | Gold: 4.8 | Water: 4.5 | Fertilizer: 3.9 | $3.12 |
| 7 | Claude Haiku 5.5Claude Code · $0.81 | Claude Code | 36.23 | Hay: 5.5 | Wood: 5.6 | Carrot: 5.3 | Pumpkin: 5.4 | Cactus: 4.2 | Power: 3.6 | Water: 4.3 | Fertilizer: 2.4 | $0.81 | |||
| 8 | GPT-6 LunaCodex · $0.08 | Codex | 15.74 | Hay: 4.0 | Wood: 2.9 | Carrot: 2.1 | Pumpkin: 2.1 | Cactus: 1.2 | Bone: 1.5 | Water: 1.9 | $0.08 |
points per resource, one per tenfold producedAll at medium effort · Cost is an estimate, not a charge · What is comparable
efficiency
Score against cost, tokens and real time
Up and to the left: more score for less
1Claude Opus 5.5Claude Opus 5.5Claude Code · mediumScore 75.5 (75.4–75.7)Cost $5.35 ($4.73–$6.19)14 points per $3 of 3 seeds2GPT-6.1 Sol74.1 · $0.47GPT-6.1 SolCodex · mediumScore 74.1 (71.8–75.9)Cost $0.47 ($0.32–$0.65)159 points per $3 of 3 seeds3Claude Fable 5.169.2 · $14.68Claude Fable 5.1Claude Code · mediumScore 69.2 (63.7–72.5)Cost $14.68 ($8.28–$21.62)4.7 points per $3 of 3 seeds4GPT-6 AstraGPT-6 AstraCodex · mediumScore 66.6 (63.2–69.9)Cost $1.79 ($1.73–$1.88)37 points per $3 of 3 seeds5Grok 4.758.4 · $11.41Grok 4.7Cursor · mediumScore 58.4 (57.4–59.8)Cost $11.41 ($9.31–$13.92)5.1 points per $3 of 3 seeds6Claude Sonnet 5.5Claude Sonnet 5.5Claude Code · mediumScore 55.1 (49.2–60.8)Cost $3.12 ($2.25–$4.38)18 points per $3 of 3 seeds7Claude Haiku 5.536.2 · $0.81Claude Haiku 5.5Claude Code · mediumScore 36.2 (34.4–37.2)Cost $0.81 ($0.61–$1.11)45 points per $3 of 3 seeds8GPT-6 Luna15.7 · $0.08GPT-6 LunaCodex · mediumScore 15.7 (4.7–27.5)Cost $0.08 ($0.02–$0.18)189 points per $3 of 3 seeds
| # | Model | Score | Cost | Seeds |
|---|---|---|---|---|
| 1 | Claude Opus 5.5 (Claude Code, medium) | 75.5 (75.4–75.7) | $5.35 ($4.73–$6.19) | 3 of 3 |
| 2 | GPT-6.1 Sol (Codex, medium) | 74.1 (71.8–75.9) | $0.47 ($0.32–$0.65) | 3 of 3 |
| 3 | Claude Fable 5.1 (Claude Code, medium) | 69.2 (63.7–72.5) | $14.68 ($8.28–$21.62) | 3 of 3 |
| 4 | GPT-6 Astra (Codex, medium) | 66.6 (63.2–69.9) | $1.79 ($1.73–$1.88) | 3 of 3 |
| 5 | Grok 4.7 (Cursor, medium) | 58.4 (57.4–59.8) | $11.41 ($9.31–$13.92) | 3 of 3 |
| 6 | Claude Sonnet 5.5 (Claude Code, medium) | 55.1 (49.2–60.8) | $3.12 ($2.25–$4.38) | 3 of 3 |
| 7 | Claude Haiku 5.5 (Claude Code, medium) | 36.2 (34.4–37.2) | $0.81 ($0.61–$1.11) | 3 of 3 |
| 8 | GPT-6 Luna (Codex, medium) | 15.7 (4.7–27.5) | $0.08 ($0.02–$0.18) | 3 of 3 |
1Claude Opus 5.5Claude Opus 5.5Claude Code · mediumScore 75.5 (75.4–75.7)Tokens 12.9m (11.7m–14.3m)3 of 3 seeds2GPT-6.1 Sol74.1 · 1.94mGPT-6.1 SolCodex · mediumScore 74.1 (71.8–75.9)Tokens 1.94m (1.26m–3.25m)3 of 3 seeds3Claude Fable 5.1Claude Fable 5.1Claude Code · mediumScore 69.2 (63.7–72.5)Tokens 16.1m (8.3m–24m)3 of 3 seeds4GPT-6 Astra66.6 · 965kGPT-6 AstraCodex · mediumScore 66.6 (63.2–69.9)Tokens 965k (924k–997k)3 of 3 seeds5Grok 4.758.4 · 20.4mGrok 4.7Cursor · mediumScore 58.4 (57.4–59.8)Tokens 20.4m (16.4m–26m)3 of 3 seeds6Claude Sonnet 5.555.1 · 10.1mClaude Sonnet 5.5Claude Code · mediumScore 55.1 (49.2–60.8)Tokens 10.1m (7.36m–13.8m)3 of 3 seeds7Claude Haiku 5.536.2 · 15.6mClaude Haiku 5.5Claude Code · mediumScore 36.2 (34.4–37.2)Tokens 15.6m (11.2m–21.2m)3 of 3 seeds8GPT-6 Luna15.7 · 5.65mGPT-6 LunaCodex · mediumScore 15.7 (4.7–27.5)Tokens 5.65m (1.21m–12.5m)3 of 3 seeds
| # | Model | Score | Tokens | Seeds |
|---|---|---|---|---|
| 1 | Claude Opus 5.5 (Claude Code, medium) | 75.5 (75.4–75.7) | 12.9m (11.7m–14.3m) | 3 of 3 |
| 2 | GPT-6.1 Sol (Codex, medium) | 74.1 (71.8–75.9) | 1.94m (1.26m–3.25m) | 3 of 3 |
| 3 | Claude Fable 5.1 (Claude Code, medium) | 69.2 (63.7–72.5) | 16.1m (8.3m–24m) | 3 of 3 |
| 4 | GPT-6 Astra (Codex, medium) | 66.6 (63.2–69.9) | 965k (924k–997k) | 3 of 3 |
| 5 | Grok 4.7 (Cursor, medium) | 58.4 (57.4–59.8) | 20.4m (16.4m–26m) | 3 of 3 |
| 6 | Claude Sonnet 5.5 (Claude Code, medium) | 55.1 (49.2–60.8) | 10.1m (7.36m–13.8m) | 3 of 3 |
| 7 | Claude Haiku 5.5 (Claude Code, medium) | 36.2 (34.4–37.2) | 15.6m (11.2m–21.2m) | 3 of 3 |
| 8 | GPT-6 Luna (Codex, medium) | 15.7 (4.7–27.5) | 5.65m (1.21m–12.5m) | 3 of 3 |
1Claude Opus 5.5Claude Opus 5.5Claude Code · mediumScore 75.5 (75.4–75.7)Real time 17m 47s (12m 32s–26m 38s)3 of 3 seeds2GPT-6.1 SolGPT-6.1 SolCodex · mediumScore 74.1 (71.8–75.9)Real time 7m 35s (6m 34s–8m 15s)3 of 3 seeds3Claude Fable 5.1Claude Fable 5.1Claude Code · mediumScore 69.2 (63.7–72.5)Real time 33m 18s (13m 52s–58m 17s)3 of 3 seeds4GPT-6 Astra66.6 · 5m 22sGPT-6 AstraCodex · mediumScore 66.6 (63.2–69.9)Real time 5m 22s (5m 01s–5m 52s)3 of 3 seeds5Grok 4.758.4 · 57m 34sGrok 4.7Cursor · mediumScore 58.4 (57.4–59.8)Real time 57m 34s (48m 41s–1h 10m)3 of 3 seeds6Claude Sonnet 5.5Claude Sonnet 5.5Claude Code · mediumScore 55.1 (49.2–60.8)Real time 13m 02s (8m 25s–18m 02s)3 of 3 seeds7Claude Haiku 5.536.2 · 11m 40sClaude Haiku 5.5Claude Code · mediumScore 36.2 (34.4–37.2)Real time 11m 40s (10m 37s–13m 10s)3 of 3 seeds8GPT-6 Luna15.7 · 9m 59sGPT-6 LunaCodex · mediumScore 15.7 (4.7–27.5)Real time 9m 59s (3m 55s–20m 23s)3 of 3 seeds
| # | Model | Score | Real time | Seeds |
|---|---|---|---|---|
| 1 | Claude Opus 5.5 (Claude Code, medium) | 75.5 (75.4–75.7) | 17m 47s (12m 32s–26m 38s) | 3 of 3 |
| 2 | GPT-6.1 Sol (Codex, medium) | 74.1 (71.8–75.9) | 7m 35s (6m 34s–8m 15s) | 3 of 3 |
| 3 | Claude Fable 5.1 (Claude Code, medium) | 69.2 (63.7–72.5) | 33m 18s (13m 52s–58m 17s) | 3 of 3 |
| 4 | GPT-6 Astra (Codex, medium) | 66.6 (63.2–69.9) | 5m 22s (5m 01s–5m 52s) | 3 of 3 |
| 5 | Grok 4.7 (Cursor, medium) | 58.4 (57.4–59.8) | 57m 34s (48m 41s–1h 10m) | 3 of 3 |
| 6 | Claude Sonnet 5.5 (Claude Code, medium) | 55.1 (49.2–60.8) | 13m 02s (8m 25s–18m 02s) | 3 of 3 |
| 7 | Claude Haiku 5.5 (Claude Code, medium) | 36.2 (34.4–37.2) | 11m 40s (10m 37s–13m 10s) | 3 of 3 |
| 8 | GPT-6 Luna (Codex, medium) | 15.7 (4.7–27.5) | 9m 59s (3m 55s–20m 23s) | 3 of 3 |
rules
How a run works
- Model + its own agent
- Claude in Claude Code, GPT in Codex, Grok in Cursor. Built-in tools are off; the only way to act is FarmerBench’s game tools.
- A paused clock
- The game waits while the model thinks. Each action costs 10 game-seconds; running a program costs the time it runs.
- Same farm, same seeds
- Seeded and deterministic, so every run can be replayed exactly and every model faces the same fields.
- Orders of magnitude
- Ten times more of anything is one more point: one more pip in the counter.
# economy-v1, over eleven resources
score = 0
for r in Resources:
score += log10(1 + produced[r])