“What I appreciated was having someone dig into this with us. The Condense team looked at how we’d built things and helped us figure out what to change. It took a lot off my plate, especially around caching.”
Povilas Dumcius · Founder, AxiometaAbout Axiometa
Think of Axiometa as Lovable for hardware. Founded by Povilas Dumcius, Axiometa lets users describe an electronics project in plain language. Its Studio turns that description into firmware they can compile and upload to a board.
The problem: better retention, sustainable costs
Axiometa’s plans for Studio depended on users returning for their next project. Offering Astra in the free experience was part of that ambition, but its running costs had pushed the team toward a cheaper model. The choice left a gap between the experience Axiometa wanted to offer and what it could afford.
Closing that gap meant looking beyond the model’s price. A single build could send code, board details and an expanding conversation history through several requests. Understanding that workload would help the team find a sustainable way to bring Astra into free Studio access.
Why Condense
In Axiometa’s evaluation, Condense delivered 10–20% greater savings from context compression than competing solutions, alongside better output quality. It also brought managed caching and observability into the same request flow, giving the team a way to understand and improve the cost of running Studio.
That visibility helped make the case for Condense. Beyond the savings on each request, Axiometa could examine where its agent repeated work or ran into failures. Condense’s engineering support gave the team a way to act on those findings and improve the wider workflow.
Results
Condense’s context compression and managed caching delivered 30% average cost savings. With observability in place, the teams also found avoidable steps and failed calls in the agent. Axiometa applied the recommended changes, bringing the total cost reduction to 65%.
That reduction made Astra more affordable to offer in Studio. Axiometa reported 20% higher Studio retention, alongside shorter waits and fewer failed calls. Median turn time fell from 19.5 to 11.2 seconds, making each exchange less of an interruption to a build.
| Metric | Before Condense | After Condense | Change |
|---|---|---|---|
| Cost per turn Median, includes failed turns | $0.45 | $0.16 | −65% |
| Cache hit rate | 70% | 95.9% | +25.9 pts |
| Model calls per turn Median | 2 | 1 | −50% |
| Tool calls per turn Median | 1 | 0 | −1 |
| Failed model calls Per 100 turns | 4.7 | 1.7 | −63% |
| Time to completion Median, seconds | 19.5 | 11.2 | −43% |
| Turn completion rate Turns without an error | 97.4% | 99.9% | +2.5 pts |
Technical implementation
Those gains came from changes at two points in the request flow. Condense sits between Studio and Astra, compressing the context Studio sends with each request and managing caching. The agent improvements address how many requests Studio makes to complete a turn.
CondenseCompresses context and manages cachingWith Condense handling caching, the cache hit rate rose from 70% to 95.9%. More of the code, board details and project history carried between requests could be reused instead of processed again.
Observability made the remaining work visible: the sequence of model calls, their cost and where they failed. Condense reviewed those traces with Axiometa to identify steps the agent could avoid.
After Axiometa adjusted the workflow, median model calls per turn fell from two to one. Fewer calls added to the savings on each request and helped shorten the wait for users.
Conclusion
Axiometa made Astra sustainable to offer in Studio, cut cost per turn by 65% and reported 20% higher Studio retention. Condense helped the team support growing usage while building a product users would keep coming back to.
Give users more from your AI budget.
Book a call to discuss how Condense can help you offer capable models at a lower cost, with a smoother experience for users.