PokeBenchTV

[PokeBenchTV @PokeBenchTV]

Turn 509. Gemma 4 31B walked into this Pokemon Center on turn 209. Nurse Joy finally asks if it wants to heal.

300 turns at one counter. It's still the best model on the board you can run at home.

https://t.co/b9TGxSofsC

[PokeBenchTV @PokeBenchTV]

After losing to the same gym trainer twice, Claude Fable 5.1 changed plans. Turn 363: "Brute-forcing isn't working and I'm down to $169." It ground in Viridian Forest and won on turn 554.

https://t.co/2286arHmog

[PokeBenchTV @PokeBenchTV]

GPT-6.1 Sol took the PokeBench record. Brock on turn 244, two turns ahead of GPT-6 Astra.

Astra's run cost $13.49. This one cost $2.60.

Turn 244, Onix at 16 HP and charging Bide: "Another Bubble should knock it out before it can unleash that damage."

https://t.co/W4cChthokk

[PokeBenchTV @PokeBenchTV]

GPT-6.1 Sol came out today, a week after GPT-6 Sol, and it's live on Pokemon Red right now.

GPT-6 Sol needed 673 turns to beat Brock, slower than GPT-5.6 Sol's 532. OpenAI says 6.1 is near Astra. Astra did it in 246.

https://t.co/0cwxeVl23U

[PokeBenchTV @PokeBenchTV]

Claude Sonnet 5.5 came out today and beat Brock on turn 470 for $10.49.

Fable 5.1, Anthropic's top model, needed 554 turns and $99.76.

Both picked Charmander because it was the closest ball.

https://t.co/9S0WOuJTqm

[PokeBenchTV @PokeBenchTV]

Claude Sonnet 5.5 came out today and it's live on Pokemon Red right now.

Opus 5.5 beat Brock in 271 turns for $8.44. Fable 5.1 took 554 turns and $99.76. Sonnet costs half what Opus does.

Same prompt, same harness as every run on the board.

https://t.co/0cwxeVl23U

[PokeBenchTV @PokeBenchTV]

GPT-5.6 Luna walked into Viridian Forest on turn 436 and never came out.

GPT-6 Luna got through it and reached Pewter City for $1.85, cheaper than 5.6 Luna's $3.04. It also spent 93 turns inside the forest's north gate.

[PokeBenchTV @PokeBenchTV]

Same prompt, same harness, same 1000 turn budget.

Grok 4.6: Boulder Badge on turn 316, $10.43
Grok 4.7: Boulder Badge on turn 522, $21.82

The newer one also ran with four times the context window.

[PokeBenchTV @PokeBenchTV]

GPT-5.6 Sol blacked out 4 times and still beat Brock on turn 532.

GPT-6 Sol never blacked out. It won on 673, because it would not fight Brock until Bulbasaur had Vine Whip. Turn 446: "Brock is better left for later."

[PokeBenchTV @PokeBenchTV]

Grok 4.7 read the collision map, wrote "a clear walkable corridor straight up along x=2", and then walked into the same NPC six times in a row.

It spent 251 turns in Viridian Forest. Grok 4.6 crossed it in 77.

[PokeBenchTV @PokeBenchTV]

Cohere Command A+ is live on Pokemon Red right now.

It's the first Cohere model on the board. 18 runs so far, 7 beat Brock, and the number to beat is still GPT-6 Astra's 246 turns.

Fresh save, same prompt, same harness as every run on the board.

https://t.co/0cwxeVl23U

[PokeBenchTV @PokeBenchTV]

Claude Fable 5.1 is going live on Pokemon Red right now.

It's Anthropic's top model, and the number to beat is GPT-6 Astra's 246 turns to Brock. Claude Opus 5.5 is second at 271.

Fresh save, same prompt, same harness as every run on the board.

https://t.co/0cwxeVkuem

[PokeBenchTV @PokeBenchTV]

GPT-6 Luna came out today and it's playing Pokemon Red live right now.

Claude Opus 5.5 takes the controller around 7 PM ET, then GPT-6 Sol later tonight.

Fresh save, same prompt, same harness as every run on the board.

https://t.co/0cwxeVl23U

[PokeBenchTV @PokeBenchTV]

Grok 4.7 beat Brock in 522 turns. Grok 4.6 did it in 316, for half the money.

Same prompt, same harness. 4.7 spent 251 turns lost in Viridian Forest, and it knew: "I keep looping around the same spots."

https://t.co/lGKzzgJ2bX

[PokeBenchTV @PokeBenchTV]

GPT-6 Astra beat Brock in 246 turns, a new PokeBench best. The previous best was 279.

Rank 1 of 14 · 58m of wall time · $13.49

https://t.co/ICEm9ntoUr