Andrej Drats

@AndrejDrats

Private X content partner for tech & health founders | Worked with multiple 9-figure founders | 600M+ views generated | $3M in revenue for clients

Apply to work with me →

Andrej Drats
@AndrejDrats
2h
which model do you actually trust to run an agent without watching it? a) opus 5.5 b) gpt-6.1 sol c) gemini 4 argon d) none, i still read every diff
1
Andrej Drats
@AndrejDrats
2h
anthropic's s-1 says they made opus 4.5 cheaper than opus 4.1, and quarterly revenue growth went from 48% to 132% after the launch please remember this when you price fable 5.5 lol
1
Andrej Drats
@AndrejDrats
2h
sonnet 5.5 is half the price of opus 5.5 per token, but in claude code it costs more per task $13.30 vs $11.42 on artificial analysis' coding agent benchmark. it also takes 1.3h vs 1h and scores lower
8
Andrej Drats
@AndrejDrats
4h
new paper tested fancy ai memory systems against the dumbest setup possible: keep the whole chat history and search it when the agent needs something the dumb one won in 6 of 7 tests. loading a 1M token chat took 0.3 hours instead of 27 try the simple version before you build
Andrej Drats
@AndrejDrats
4h
a chinese job site (boss zhipin) just put out a free 33B model, and it's better at running a startup than qwen or gemma lol yc-bench hands the model $200k and a fake ai company for one simulated year. it ended with $1.02M, qwen 3.8 with $641k, gemma 4 with $265k