Show HN: Echo – Fable-level results at 1/3 the cost using open-weight models
A new service called Echo claims to match the performance of Anthropic’s high-end Fable model at roughly one-third the cost by routing queries across multiple open-weight language models. Commenters are intrigued by the potential of model routers and ensemble-style systems to cut inference costs, but question the lack of transparency around which models are used, limited and sometimes weak benchmarks, and UI choices they view as dark patterns. Many see value in intelligent orchestration of different models for different tasks, while warning that real-world savings and reliability will depend heavily on workload, observability, and how routing is implemented.
Product concept & claims
- Echo is presented as a router/orchestrator over open‑weight LLMs, aiming for “Fable‑level” quality at ~1/3 cost.
- It decides which models to use and how much compute a request merits, sometimes combining intermediate work rather than always running a full ensemble.
- It’s explicitly not just a single new model; it relies on dynamic routing and allocation across a pool of open‑weight models.
Benchmarks and technical approach
- An eval dashboard exists with ~900 rows across several benchmark families, but multiple commenters find these benchmarks weak, incomplete, or misaligned with the “Fable‑level” claim, especially for coding/SWE.
- Discussion notes that some benchmarks (e.g., GPQA Diamond) may be saturated or contain label errors, making small improvements difficult or misleading.
- There is interest in more rigorous, harder benchmarks (coding, agentic tasks) and comparisons to router‑focused benchmarks like RouterBench.
- Some argue routing can cut cost but cannot make the system fundamentally “smarter” than its best underlying model.
Trust, transparency, and UX concerns
- Echo does not reveal which specific model handled a given request, calling the routing policy its core IP.
- Multiple participants say this is a deal‑breaker for serious or enterprise use due to debugging, observability, compliance, and capacity‑planning needs.
- The site initially had a fake “chat” box that redirected to signup and password‑complexity constraints; this is criticized as a dark pattern.
- Signup friction (login errors, “too many authentication attempts”) and privacy‑policy wording also drew criticism; the founder says these have been or are being fixed and that prompts/outputs are not used for training.
Comparisons to other routers / MoE systems
- Frequently compared to OpenRouter Fusion, Sakana Fugu, and other “model routers.”
- Fusion‑style systems run multiple models then synthesize, improving quality at higher cost/latency; Echo aims for similar gains with lower cost/latency via selective routing/ensembling.
- Parallels are drawn to classic ensemble methods and mixture‑of‑experts architectures, though Echo operates per‑request (not per‑token).
Cost, pricing, and who benefits
- Commenters note savings depend heavily on workload: if many queries are easy, routing to cheaper models can help; otherwise, benefits shrink or disappear, especially once caching is considered.
- A long subthread debates whether this matters for users on heavily subsidized frontier‑model subscriptions (e.g., $100–$200/month plans) versus API/enterprise users paying list token prices.
- Some argue open‑weight routing is more about future, less‑subsidized economics and about regions/segments where high subscription prices are prohibitive.
Broader reflections
- Several see multi‑model orchestrators as a likely end‑state: “the best model” becoming niche, with most systems using routers that juggle cheap and strong models.
- Others are skeptical, viewing many such projects as crowded, incremental, or primarily cost‑optimization rather than fundamentally new capability.