OpenAI case study: how Amplitude's Wave agent ships product improvements in 1-2 days

OpenAI published a customer story on November 12 about Amplitude, the product analytics company, and its proactive product agent Wave. According to the account, roughly 80% of Wave runs on OpenAI models, including GPT-6.1 Sol, GPT-6 Luna and GPT-Image-2.5 Flare.

Wave is not positioned as an idea generator but as a path from product data to shipped change. A team hands it a product area plus the goals and metrics it wants to move, and Wave pulls together Amplitude data, session replays, customer feedback and code to identify and rank opportunities, draft a solution, and coordinate implementation with people and coding agents. When an opportunity is ready to build, the plan and acceptance criteria go straight to agents such as Codex, and Wave then tracks the outcome.

The models are split by task. GPT-6.1 Sol supports the opportunity manager and deep investigations into code and data, GPT-6 Luna handles high-frequency, lower-cost work such as scanning code and running small evaluation checks, and GPT-Image-2.5 Flare generates product mockups using screenshots from session replays as visual references.

The most concrete example is Amplitude's own homepage. In a sample of 100 session replays, 80 users typed their own questions while only 12 selected a generic suggestion card. Wave proposed personalizing the cards based on recent charts, dashboards or the referring page, keeping free-form typing intact and falling back to generic suggestions when personalization context was missing, then handed the change to a coding agent for an A/B test. One day later, messages sent through the cards were up 160% against control.

Amplitude also published model evaluations. On the Opportunity Investigator task GPT-6.1 Sol scored 4.06 out of 5 against 3.62, 3.52 and 2.94 for three alternatives, led on all five investigation criteria and had the fastest median investigation time at 54 minutes. On the Data Science Agent it scored 67.5 versus 69.3 for another leading model, at a mean cost of $1.28 per dispatch against $2.63, and its image model scored 80.3 for solution wireframes against 63.9.

Wave is in beta with customers across industries, and teams choose how much automation they want: most send bug fixes straight to coding agents, while new features and bigger bets keep a human review. Among customer-launched experiments, flags and guides, one in four went live the same day Wave surfaced the opportunity, half within four days and two in three within a week.

This is a vendor case study, so the productivity and quality numbers are Amplitude's own and cannot be independently checked. The direction it points to is still notable: the real change is not a stronger model but a loop that runs from spotting an opportunity to writing the plan, handing it to a coding agent and measuring the result in a day or two. Amplitude CEO Spenser Skates frames the ambition as making self-improving products the default way companies build software; the open question is whether the loop reproduces as reliably outside the beta.