private preview
Replace frontier-model spend on the workflows your team already understands.
Understudy is in private preview with a small group of design partners. We're shipping the first replacement model on their production workloads - proxy installed, traces captured, training end-to-end, deployed in their stack. If you're paying frontier prices for work your domain experts already know, this page is for you.
The preview is a committed partnership. We pick a real workflow, we ship a working replacement model, and we figure out together what we needed to build to make it stick.
Teams running production LLM workflows where the cost or latency is starting to hurt - and where the people who know what good looks like aren't on an ML team.
You're a fit if
- You spend meaningfully on frontier APIs (Anthropic, OpenAI, or both) on a workload you can identify and isolate.
- You have at least one product surface where domain experts (PMs, ops, support, sales) review or shape model outputs today.
- You have an engineer who can install a proxy in your coding agent or production environment in a 30-minute session.
- You can commit a domain expert to roughly 2 hours/week of evaluation work for the duration of the engagement.
Later-stage fit
- You've already built internal post-training infrastructure with dedicated ML talent.
- Your LLM work is still exploratory, without real traffic or a stable task to isolate.
- Your data has to stay inside a single approved vendor's cloud. We can work toward this after the early-preview workflow matures.
- Direct access to both founders. We're in the workbench with you throughout the engagement.
- A working open-weight replacement model - trained on your traces, evaluated against your held-out test, served on your infrastructure.
- Hands-on partnership pricing during preview. We'd rather earn your business than bill for the early work; specifics in a call.
- Shared learnings from the cohort, anonymized. You see what's working at other teams before it ships.
- Production trace access for at least one workload, through offline export or live capture via our open-source proxy.
- A domain expert available for roughly 2 hours/week of evaluation in the workbench.
- Honest feedback: what's working, what's broken, what's missing.
- The right to anonymize and cite the work in our own reporting. No logos without your written approval.