The premise of Echo, from YC-backed TracerML, is simple to state and surprisingly hard to pull off: expose one OpenAI-compatible endpoint, but behind it coordinate a fleet of open-weight models so that each request lands on whichever combination of models will handle it best. The result, they claim, is Claude Fable-level quality at roughly one-third the inference cost.
The product went into public alpha today, with an evaluation observatory showing 907 questions across eight benchmarks. On several of those, Echo matches Fable. On others — Belebele, Global-MMLU, MMLU-Pro specifically — Fable still leads. That’s the honest version of the story, and to TracerML’s credit they’re publishing it question-by-question rather than cherry-picking aggregate numbers.
The technical foundation traces back to their April routing paper (TRACER), which trains lightweight surrogates on production LLM logs to handle classification tasks without touching the full model. The system uses a parity gate: the surrogate takes over only when its outputs agree with the teacher LLM above a configurable threshold. For tasks where the surrogate can’t be trusted, it defers to the full model. Echo appears to generalize this idea to generation tasks — routing based on where each open-weight model is reliably strong, not just where the frontier model is overkill.
The key insight from the creator, Adam Rida, is worth sitting with: “One thing that surprised me while building it was how complementary the models are. A model that is clearly weaker overall can still be extremely useful on particular problems.” This is different from the standard cost-quality tradeoff framing, where you pick a cheaper model when quality requirements are lower. It suggests that routing is better modeled as task-model affinity than as capability tiering — that the right model for a given query isn’t always the most capable one, even if you could afford it.
Whether that holds at scale across diverse production workloads remains to be seen. TracerML isn’t disclosing which open-weight models are behind Echo or the per-model routing logic, positioning that as core IP. That makes independent verification tricky: you’re trusting their eval methodology and the claim that “open-weight models” means something consistent with what’s publicly available. Their evaluations are self-published rather than third-party audited, which they acknowledge directly.
Still, the results on several benchmarks are plausible. Open-weight models have advanced quickly — Laguna S 2.1 from Poolside (reviewed earlier this month) hits 70.2% on Terminal-Bench 2.1 as a single 118B-active MoE. DeepSeek V4 Pro holds 80.6% on SWE-bench Verified. The models are genuinely strong. If they’re complementary in the way TracerML claims, an ensemble that routes well could plausibly close most of the remaining gap to proprietary frontier models.
The pricing model makes this interesting for small teams and startups that can’t absorb Fable’s API costs. Echo is currently free in public alpha with invite codes (10 per account). What happens when it exits alpha and pricing materializes is a different question — the one-third cost claim needs a denominator, and if it’s priced against Fable at retail rather than volume rates, the math changes for everyone except the smallest callers.
For developers who’ve been watching the open-weight ecosystem get stronger every quarter and wondering when it becomes viable to drop proprietary model subscriptions entirely: Echo is a concrete bet on that thesis, and they’re backing it with evals rather than marketing copy.