Many production workflows run one narrow task millions of times on a frontier model. A small model trained for that task can cut the cost sharply, and it no longer has to take a quarter of ML engineering time.
Frontier prices for a narrow job.
A team points a frontier model at one repetitive job, like classifying messages or choosing the next tool, and runs it millions of times a month at frontier prices. When a small open model is eventually fine-tuned for that job, the cost collapses. Orchestra cofounder and CEO Luis Manrique has seen one workflow drop from $400K a year to a few hundred dollars a month.
The catch is engineering time.
The last such project he saw consumed more than a fiscal quarter of a top engineer's time. Savings near 100x demand the scarcest, costliest people on staff, and most of these projects never get staffed.
The stakes are high. ICONIQ's 2026 State of AI puts the average AI product near a 52% gross margin, compared with more than 80% for legacy SaaS, and inference alone takes close to a quarter of revenue. Cursor, Ramp and Intercom already build custom models to replace off-the-shelf ones, but they have deep in-house teams. Orchestra aims to make this a one-week job instead of a one-quarter job.
How Orchestra works.
It runs inside the coding agents your team already uses: Claude Code, Codex, Cursor and any MCP-compatible tool.
It records traces as agents work, spots the best optimization targets for a set goal, and compares techniques.
It runs locally on commodity hardware, keeping data in your environment until a task justifies the cloud.
It steps from prompt tuning to supervised fine-tuning to reinforcement learning as each task earns it.
You own the prompts and weights, and serve them on your GPUs, Fireworks, Bedrock, Vertex or elsewhere.
A small share of production traffic stays in ongoing experiments, so the model keeps improving.
One design partner was spending five figures a month on tasks that never required a frontier model. On a model it now owns, it cut that spend by more than half. Orchestra publishes its benchmarks, unflattering ones included.
Who it is built for.
Orchestra's method for judging whether one model beats another at a job began with Luis and his cofounder at Instacart. It is built for teams feeling cost or latency pain in production, and for teams whose quality judges are not model trainers. The full essay has the founder's first-person account.
Evaluate your workflows and see the published results for yourself.