Inside the doja model router — picking the right brain per job.
No single model wins everything anymore. The smartest AI-native systems treat frontier models the way a studio treats musicians — different player, different track, one tight arrangement.
Through 2026 the frontier stopped being a leaderboard and became a portfolio. We rewired doja so every task is dispatched to the model that is currently best at it, then prompted with the tightest possible instruction set — which is where the real cost savings come from.
The top of the bench
Eureka leans on Claude Opus 4.8 for the coding and design execution, with GPT-5.5 transforming raw briefs into clean specs and Gemini 3.5 Pro pulling in grounded context when a build touches a regulated or research-heavy domain. Limitless lets Gemini 3.5 Pro drive stakeholder research and grounding, GPT-5.5 handle the quality analytics on every message variant, and Claude Opus 4.8 design the sequences themselves. Gatsby grounds every battlecard in Gemini 3.5 Pro deep research before GPT-5.5 transforms it into the final brief.
The model is no longer the product. The orchestration — and the prompt — is.
Accuracy is the cheapest optimisation. A sharper prompt against the right model means fewer retries, fewer tokens, fewer fallbacks — which is how we keep the 100× speed and 10× cost promise even as new models ship every month.
Tell us what you're building.
Drop us a note — founders, operators, enterprises, press. The team reads every message and replies inside a day.