FRAMEWIREIndonesiaUpdated Sep 12Live wire
0:00 / 0:00

Towards AI spent about $600 on evals racing 6 context-trimming techniques against doing nothing

Doing nothing won Keeping the full history cost $0.11 per turn against $0.24 for their tuned production setup, answered 4 seconds faster, and recalled 92% of earlier facts against

TreffSep 125
0:00 / 0:00

Deepseek-v4.1-flash made this, with help from LTX 2.5

⛏️🌱Don Kackman 🌱⛏️Sep 122
0:00 / 0:00

DeepSeek-V4.1-Flash

Fabiano FirmoSep 121
0:00 / 0:00

First look at today's value board: GLM 5.3 Flash leads with a 42 score at just 9 cents

While DeepSeek V4 Flash offers the cheapest entry at 3 cents.

KFChow AI LabSep 12