Developers use Jev router to split coding tasks across five AI models
A developer said TypeSafe's Jev model picked which of five frontier models handled each part of a coding demo, without him touching a model selector, cutting cost 40%.
The developer known as omarsar0 said he built a full-stack demo app inside a single Codex session. TypeSafe's Jev router, rather than the developer, assigned each step to a different frontier model. Claude Opus 5.5 planned it, GPT-6 Astra built the backend, and Kimi K3 built the frontend.
https://x.com/omarsar0/status/2104623036133417456
DeepSeek handled testing and GLM 5.3 Flash wrote the documentation, omarsar0 said. He added that he never touched a model picker, and the build came in 40% cheaper and 15% faster than his usual workflow. Separately, the commentator kimmonismus described the same routing mechanism. Kimmonismus said swarms of Jev instances score each part of a conversation against benchmark results, then vote on which model takes each step.
Kimmonismus put the result at 99% of Claude Opus 5.5 and GPT-6 Astra's own performance on three benchmarks: DeepSWE, Terminal-Bench 4.0 and SWE-Atlas. The savings came to 40% less cost, he said. TypeSafe has not published the figures itself, and neither account's numbers have been independently verified.
Kimmonismus said a router that reaches near-Opus results, while only calling the expensive model when needed, could make premium model tiers look optional to some customers. Claude Opus 5.5 separately entered LMArena's Agent Arena leaderboard at No. 2 this week, behind only Fable 5.1 Max. Arena put the cost at 40% to 56% lower than Anthropic's own prior Opus 5 variants.