The deep dives.

Reads where questions get run to ground
with real data.

Lost in optimization

Does squeezing a model into fewer bits hurt a coding agent? We ran FP8, INT4 and NVFP4 through multi-turn agentic benchmarks.

We had two frontier lab coding agents face the same 64 long-horizon coding tasks and ran the numbers on cost and capability (and honesty).

Your coding benchmark scores are probably wrong: your eval environment is part of your score.

The read + the takeaway

Your harness and model combination can double your token bill.

Watch the
invoice.

Your harness and model combination decides whether one resolved task costs $0.35 or $0.70.

Doom-looping agents and prefix-caching errors land straight on your bill, so keep an eye on your telemetry.

Earlier this year

A back-of-the-napkin guide to why one chat barely registers and a few agentic sessions make your GPU sweat.

A reality check on self-hosting the model behind your coding agent.