Tracer, the San Francisco AI lab founded by Adam Rida, launched Echo on July 31st, an adaptive inference system that coordinates multiple open-weight models and returns a single answer through an OpenAI-compatible API. Tracer said in an eight-post thread on X that Echo approached the quality of Anthropic's Claude Fable 5 across its current evaluations at roughly one-third of the inference cost.
https://x.com/tracerml/status/2083297340207886775?s=46 Rida is building Tracer alone. Y Combinator lists Tracer as a one-person Summer 2026 company based in San Francisco. Rida previously studied machine learning interpretability as a PhD candidate at Sorbonne University and CNRS, then built applied ML systems in finance, insurance and supply chain. In Tracer's YC launch post, Rida said he grew his previous product, DeepRecall, to about EUR100,000 in annual recurring revenue as a solo founder in under three months.
Echo extends Rida's earlier work on deciding when an application needs an expensive model call. His April 16th research paper introduced TRACER, an open-source system that learned from production traces and sent predictable classification requests to lightweight machine-learning models while deferring uncertain inputs to a larger language model.
Echo applies that cost-allocation thesis to general language-model workloads. For each prompt, Tracer says Echo decides how much compute to use, selects models from a pool of open-weight systems and combines their outputs. Developers retain their existing OpenAI client and change the base URL, model name and API key. Tracer is pitching Echo for chat, code and agent workloads, including use through OpenCode and compatible agent harnesses.
Tracer's benchmark claim
Tracer reported that Echo scored 98.6% on MATH-500 at a recorded inference cost of $4.58. Claude Fable 5, which Anthropic prices at $10 per million input tokens and $50 per million output tokens, scored 99.8% at $12.64 in Tracer's run.
On LiveCodeBench, Echo scored 92.8% at $5.35, compared with Fable's 92.0% at $8.71, according to Tracer. Those results put Echo slightly ahead on that coding test while using less recorded inference spend.
The comparisons remain Tracer-run benchmarks, rather than independent evaluations. Tracer's Echo Eval Observatory lists 907 questions across eight benchmarks and nine test sets. It publishes stored questions, answers and grades for its completed MATH-500 cohort and warns that model outputs can change between runs. Tracer also acknowledges that public benchmarks may appear in model training data and cannot establish that Echo will outperform Fable across production workloads.
The limits are material. Tracer says Echo trails Fable in multilingual tasks and broader knowledge, while LiveCodeBench does not measure repository-scale software work or the longer agent loops that require planning, tool use, code changes and error recovery. Tracer is running those longer-horizon tests, with SWE-bench Verified, ARC-AGI and BigCodeBench listed without results on the evaluation site.
Coordination instead of a single route
Model routing is already a contested infrastructure category. Not Diamond sells routing and prompt optimization, while the open-source RouteLLM framework directs easier requests to cheaper models and reserves stronger systems for harder prompts.
Tracer's technical and commercial bet goes further than choosing one model for each request. Echo can invoke multiple open-weight models, vary the compute budget and assemble their work into one response. That approach may improve answer quality, though every additional model attempt consumes tokens and adds latency. Tracer must prove that its coordination gains consistently outweigh that overhead on real application traffic.
The product is available through the Echo chat and API interface. The browser interface offers five guest messages without an account, while its account and API screens label the service a private alpha. Tracer says public Echo does not train on private user traces.
Tracer is also using the launch to recruit AI teams with substantial production inference bills. Rida says Echo can be configured around a customer's traffic distribution, quality threshold and data requirements, allowing Tracer to learn which requests merit more computation. That workload-specific service is the clearest route from a public model demo to a business: sell lower inference spend to AI applications whose gross margins shrink every time an agent takes another model turn.
Y Combinator is Tracer's only identified backer. YC's published standard deal invests $500,000 in each accepted batch company: $125,000 for 7% and $375,000 through an uncapped most-favored-nation SAFE. For Rida, that financing supports a narrow bet with broad implications for AI applications: premium model capability may increasingly be delivered by software that allocates and combines cheaper models, rather than by sending every request to the most expensive endpoint.