Amplitude's Wave agent runs mostly on OpenAI models
Amplitude says its Wave product agent, about 80% powered by OpenAI models, cuts product-improvement cycles from roughly a month to one or two days, routing tasks between models by cost and capability.
Amplitude's proactive product agent, Wave, is built primarily on OpenAI models, according to an OpenAI customer case study. Teams give Wave a product area and target metrics. It combines analytics data, session replays, customer feedback and code to find improvement opportunities, writes implementation plans for coding agents such as Codex, and then measures the results.
Amplitude estimates Wave cuts the time to ship a product improvement from roughly a month to 1–2 days. On Amplitude's own homepage, Wave noticed that users skipped generic suggestion cards: in 100 sampled session replays, 80 users typed their own questions and 12 picked a suggestion. Wave proposed personalizing the cards. A coding agent built the change, and in an A/B test, messages sent through the cards rose 160% against control after one day.
Amplitude splits the work across models. GPT‑6.1 Sol handles deep investigations, GPT‑6 Luna runs frequent low-cost checks, and GPT‑Image‑2.5 Flare generates mockups. In Amplitude's internal evaluations:
- Opportunity Investigator: Sol scored 4.06/5, ahead of three alternatives (3.62, 3.52 and 2.94), with a median investigation time of 54 minutes.
- Data Science Agent: Sol scored 67.5/100, slightly below another leading model's 69.3, at about half the cost per dispatch ($1.28 vs $2.63).
- Wireframes: OpenAI's image model scored 80.3, against 63.9 for a competing image model.
The competing models are not named. All figures come from Amplitude and were published by OpenAI. Wave is in beta with customers, and most send bug fixes straight to coding agents while larger features get human review.