Show HN: Echo – Fable-level results at 1/3 the cost using open-weight models

Show HN introduced Echo, an adaptive model assembled entirely from a pool of open-weight models that includes GLM-5.2, Kimi K2.7, and many others. The project began from a measured result: a perfect oracle over that pool beat the state of the art on every benchmark the author tested. Echo is the practical system built to capture as much of that oracle gap as possible without assuming perfect routing.

Architecturally, Echo is one adaptive model over the pool rather than a single fixed open-weight checkpoint. Routing and selection decide which open-weight model handles each request so the system can specialize by task. On the author’s first internal task mix, that setup consistently beat the best individual model in the pool and reached Fable-level results at roughly one-third the total inference cost.

For engineers and builders, the signal is cost and capability without locking into one closed API. If a pool of open-weight models can match Fable-level quality at about one-third the inference spend, teams that already run or can host open weights have a concrete path to lower unit cost while keeping quality competitive with a strong proprietary baseline.

Competitive context sits between single open-weight models and proprietary systems such as Fable, with Claude and Codex as the broader high-end comparison set on HN. The oracle result implies the ceiling of the pool is above any one model in it, and above the prior SOTA on the author’s benchmarks. Echo’s contribution is turning that ceiling into a usable adaptive product instead of leaving the gain on the table as a theoretical upper bound.

What to watch next is whether the internal-task-mix gains hold on public benchmarks and broader workloads, how much of the perfect-oracle margin Echo actually captures in production routing, and whether the one-third cost figure stays stable as the pool and traffic mix change. The author notes the project is still early, so the main follow-up is reproducibility outside the first internal mix and clarity on failure modes when the wrong open-weight model is chosen.