
10,000 Skittles sorted one at a time: Jev vs Claude Sonnet 5
10,000 Skittles, sorted by colour one at a time.
OpenJev vs Laya. Same board, same food, same decisions. One ate 47. The other ate 4.
Two snakes on the same 20 x 20 board, chasing the same food in the same order, for the same 30 seconds. Every move on both boards is a live call to Jev, TypeSafe's decision model, over OpenRouter. The only thing I changed between the two is how fast the answer comes back.
The snake moves one cell each time an answer lands. So the speed on the board is the decision rate and nothing else. There is no game clock I picked that could tilt the result.
On the left, every answer is applied at 38.4ms, which is Laya's published latency for one question on a T4 GPU. On the right, every answer lands at its real measured round trip, 376ms at the median. The fast lane made 782 moves and ate 47. The slow lane made 78 moves and ate 4.
The honest part: I did not run Laya. It has no hosted API anywhere I could find, and my laptop is not the hardware its numbers were measured on. So the left board is Jev's decisions at Laya's published speed. It shows what decisions per second buy you, not how Laya itself plays snake. The caption said so and so does the frame.
The whole run was 860 live calls and cost $0.017.
| Entrant | Tool | Time | Cost | Verdict |
|---|---|---|---|---|
| Laya lane (Jev's decisions) | Laya's published latency | 38.4ms per move | $0.015 | 47 food, 782 moves, 1 death |
| Jev | OpenRouter, System One | 376ms per move | $0.0015 | 4 food, 78 moves, 0 deaths |
The reels are the test bench. The real work is picking the one job in your business worth handing to AI, building it properly, and keeping it running. That’s what I do as your AI partner.

10,000 Skittles, sorted by colour one at a time.
Open
Four Claudes. One airport.
Open
One brief to two AI agents: build a 3D Rick Sanchez out of nothing but code.
Open