curated · 2026-09-20

The numbers are what make this worth a click: GPT-6 Astra alone scores 74.1% at $6.52 per task. An oracle router across 18 models hits 97.6% at $1.88. Open-weight models alone reach 90.3% at $1.45. I'd discount the 97.6% figure. It's an oracle score—they ran all models first, then picked the winner in hindsight. FireRouter has to make that choice before the task runs, and the post admits that's the hard part without giving any technical detail, latency numbers, or offline benchmarks. The useful bit is elsewhere: 94 of 113 tasks need a model under $3, and the three priciest models are the best pick on only three tasks. Most coding tasks don't need a flagship model. The router's value is putting cheap models on the right problems. Whether FireRouter can do that reliably is still just a title and a promise.