What happened
Productivity software company Asana has cut its model costs by 76 times in recent browser agent tests. The results were published in a case study by OpenAI. Asana achieved the cost reduction by using a model identified as GPT-6.1 Sol for specific tasks within its platform. The study focuses on the efficiency gains that come from strategic model selection for building scalable AI products.
How the room's reading it
The 76x figure is turning heads among developers building AI agents. On X, the immediate reaction focuses on the economic viability of complex agentic workflows, which have often been prohibitively expensive. Infra teams are pointing to this as a win for multi-model systems, where smaller, specialised models handle routine tasks instead of a single, monolithic API call. However, some practitioners are asking for more detail on the performance trade-offs. The consensus is that this highlights a shift from chasing raw capability to optimising for cost and latency — a sign of a maturing product ecosystem.
Sailfish's take
We're not surprised by the number — we're encouraged by the methodology. For months, the default for building agents has been to throw the most powerful model at the problem and wear the cost. This case study confirms what we've seen in our own builds: that approach doesn't scale. Profitability for AI products won't come from a single magical model, but from smart routing and optimisation. It’s about using the right tool for the job. We see this as a clear signal. If you're still using a flagship model for every single user interaction, you're burning money without a good reason.