Token efficiency is the number I actually optimize for — resolution rate matters, but a harness that burns 30% fewer tokens on equivalent tasks gives you more iterations and richer context before hitting limits. Would be interested to see this breakdown per-model rather than per-harness; same harness with different models can have very different efficiency profiles.

There's no point in comparing that. As long as you use a model and harness from the same company, it will always be cheaper than relying on middlemen like Microsoft. But the more expensive AI gets, the more people will do Bring Your Own Key or host their own models. In that case, the efficiency of the harness is very important to know.

Periodic Neon is an early example: using our lab data, we trained a trillion-parameter model that outperforms GPT-6 Astra on a critical scientific analysis task using relatively little compute.

Eight out of eight models detected Mythos's flagship FreeBSD exploit, including one with only 3.6 billion active parameters costing $0.11 per million tokens.

In 23 months, the same capability that needed 1.8 trillion parameters now fits in 4 billion parameters. A 450x compression.