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

adam_rida posted Echo on Hacker News as a Show HN, where it reached 484 points and 229 comments. Echo is live at https://echo.tracerml.ai/ and is framed as an experiment in building one AI system from a pool of open-weight models instead of picking a single model for every task. The headline claim is Fable-level results at one-third the cost by relying on open-weight models rather than a fixed proprietary stack.

The technical starting point was a controlled comparison: a group of models, including GLM-5.2, Kimi K2.7, and others, was run on the same evaluations. The author then measured an oracle-style upper bound—what performance would look like if, for each problem, you already knew which models would help and how their outputs should be combined. That hypothetical system beat any individual model in the pool by a substantial margin. The catch is deployability: the oracle depends on knowing which decisions were good after the fact, so it is a ceiling on ensemble potential, not a shippable routing policy by itself.

For engineers and builders, the useful signal is not “more models always win.” It is that evaluation-driven pool design can expose complementary strengths that a single open-weight choice leaves on the table. If your product already spans coding, reasoning, and mixed task types, treating the model layer as a selectable pool with a measured combine strategy may matter more than chasing the next single-model leaderboard jump—provided you can turn the oracle insight into something that works without hindsight.

Competitively, Echo sits in the growing middle ground between one-model APIs and full custom infra. The Show HN pitch competes on cost and quality together: match a high bar associated with Fable while paying about one-third, using open weights. That framing pressures both closed high-end providers on unit economics and single open-weight deployments on whether one checkpoint is enough when several strong open models already exist in the same evaluation set.

What to watch next is whether Echo can close the gap between the oracle experiment and a real production system—routing and combination that do not require knowing good decisions in advance—while keeping the stated cost advantage. The evaluation setup already implies a practical checklist: define the task mix, measure each pool member on the same suite, quantify the multi-model ceiling, then only ship a combiner that recovers a meaningful share of that ceiling without exploding latency or spend.