Combining open-source AI models for better answers at lower cost
Hi! I’m Adam, founder of Tracer.
We built Echo, an adaptive AI system built entirely from open-weight models. Instead of making you choose between models, Echo decides how much compute each request needs, which models should work on it, and how their answers should be combined.
Across our current eval mix, Echo reaches Fable-level performance at roughly 1/3 the inference cost.
Today, Echo is strongest on math, science, and the coding benchmarks we currently cover. Multilingual and broader knowledge performance still has room to improve, while longer-horizon coding and agentic evaluations are underway.
Echo is continuously updated as we expand the model pool and improve ensembling and cross-model coordination.
Why we built Echo
There are now dozens of strong models, each with different strengths, weaknesses, and prices.
We started with a simple experiment: what if you had a perfect oracle that always knew which model would perform best for each request?
The oracle was substantially stronger than any individual model in the pool.
Echo is our attempt to recover that advantage without knowing the answer beforehand.
Instead of picking one model for everything, Echo goes beyond routing and dynamically allocates compute across a pool of open-weight models behind a single interface.
The result
We evaluate Echo against frontier models across reasoning, math, coding, multilingual and other benchmarks.
We measure both quality and the actual model inference cost used to produce the answers.
On MATH-500, Echo reached 98.6% at $4.58, compared with Claude Fable at 99.8% at $12.64.
On LiveCodeBench, Echo reached 92.8% at $5.35, compared with Claude Fable at 92.0% at $8.71.
These results show where Echo is already strongest. LiveCodeBench is encouraging, but it does not capture the full difficulty of repository-level and agentic coding. We are currently running longer-horizon evaluations where Echo must plan, use tools, modify code, and recover from errors across multiple steps.
Those results are visible in the Eval Observatory and are active areas of improvement.
We publish the underlying evaluations, including individual answers, costs, wins, and losses:
→ Explore the Echo Eval Observatory
From open source to Echo
We first explored this idea through Tracer's open-source work: learning from real AI workloads where expensive model calls are actually necessary and where cheaper paths are enough.
The project has now crossed 1,000 GitHub stars, with developers using and experimenting with the approach in their own AI pipelines.
Echo is the next step. Instead of optimizing individual calls with a routing layer, we're building one adaptive system that decides how intelligence should be allocated end to end.
→ GitHub
Every company’s AI traffic has a shape
Our public evaluations test Echo across a broad mix of general tasks. Production workloads are different.
Every company has its own traffic distribution and quality requirements. Some requests genuinely need expensive frontier-level intelligence. Many don’t.
For teams that work with us, Echo can be configured to learn the structure of their workload and identify where expensive intelligence is useful and where cheaper paths are enough.
Public Echo does not train on users’ private traces. Workload-specific learning is enabled only for teams that explicitly work with Tracer to configure it.
We set up this optimization with each team around its workload, quality bar, and data requirements.
The goal is simple: an AI system that continuously learns how to spend the minimum amount of inference required to maintain the quality you need.
That means ~3x cheaper is not necessarily the ceiling for a real production workload.
Try Echo now
Echo is publicly available through a chat UI and an OpenAI-compatible API.
You can try it directly, connect an existing application by changing the base URL, or plug Echo into OpenCode and other agent harnesses.
Echo is continuously updated as we improve model selection and ensembling.
We’re looking for teams with large inference workloads
If your company is running meaningful LLM inference in production, we’d like to test Echo against your workload.
We’ll map the quality-cost frontier on your traffic and determine how much of your inference actually requires expensive intelligence.
Teams can then work with us to activate private, workload-specific optimization.
→ Optimize Echo for your workload
About Tracer
I’m Adam Rida, founder of Tracer. I was previously a PhD candidate at Sorbonne University/CNRS, where I conducted and published research in ML interpretability. Before founding Tracer, I built applied ML systems across finance, insurance, and supply chain, then grew DeepRecall to ~€100k ARR as a solo founder in less than 3 months.
I’ve spent most of my career studying how models make decisions and turning that research into production systems.
