Amplitude says its Wave agent cuts product shipping to 1–2 days, using OpenAI models

OpenAI published a customer story on 12 November 2026 about Amplitude's Wave, a proactive product agent that finds improvement opportunities, passes plans to coding agents such as Codex, and measures the results. Amplitude estimates the cycle takes 1–2 days instead of roughly a month.

What happened

Amplitude, a product analytics company, built Wave to shorten the path from spotting a problem in usage data to shipping a fix. According to OpenAI's account, a team gives Wave a product area plus the goals and metrics it should improve. Wave then combines analytics, session replays, customer feedback and source code to rank opportunities, propose solutions and coordinate work between people and coding agents. After launch it tracks outcomes and feeds them into the next round.

Amplitude describes Wave as roughly 80% powered by OpenAI models. The story says Wave is in beta with customers across several industries.

Jacob Newman, Principal AI Product Manager at Amplitude, summarizes the idea this way:

"The idea with Wave is you have a unified platform with agents that comb through massive amounts of session replays, analytics, and code to find opportunities to improve products"

How the work is split across models

Model	Role in Wave, per Amplitude
GPT-6.1 Sol	Powers the opportunity manager and in-depth code and data investigations
GPT-6 Luna	High-volume, lower-cost work such as code scans and small evaluation checks
GPT-Image-2.5 Flare	Product mockups, using session-replay screenshots as visual references

Amplitude says it evaluates quality, latency and cost for each task before assigning a model. CEO Spenser Skates puts the benefit like this:

"OpenAI gave us both strong reasoning models and smaller, faster models, so we could balance quality, latency, and inference cost across the product"

The homepage experiment

Amplitude tested Wave on its own homepage, where generic suggestion cards were meant to prompt users to start a conversation with an AI agent. In a sample of 100 session replays, 80 users typed their own question and only 12 picked a suggestion.

Wave proposed personalizing the cards using a user's recent charts, dashboards, or the page they came from. Its technical plan kept free typing and fell back to generic cards when no personalization context existed. After a short review, the team handed the work to a coding agent and ran an A/B test.

Amplitude reports that messages sent through the suggestion cards rose 160% relative to control after one day. That is one experiment on one surface, and the story gives no baseline volumes.

Key details

• Speed: Amplitude estimates 1–2 days to ship an improvement with Wave, down from roughly a month.
• Opportunity Investigator evaluation: GPT-6.1 Sol scored 4.06 out of 5 overall, against 3.62, 3.52 and 2.94 for three unnamed alternatives. Amplitude says it led on all five criteria and had the fastest median investigation time, 54 minutes. Model cost for this evaluation was $1.34.
• Data Science Agent evaluation: GPT-6.1 Sol scored 67.5 out of 100 versus 69.3 for another leading model, so it did not top the overall score. Mean model cost per dispatch was $1.28 versus $2.63, which Amplitude reports as 51% lower. It scored higher on tool-contract compliance, estimation and uncertainty, and evidence grounding.
• Wireframes: OpenAI's image model scored 80.3 out of 100 versus 63.9 for another image model, across 15 opportunities.
• Launch timing: Among customer-launched experiments, flags and guides, one in four went live the same day Wave surfaced the opportunity, half in under four days, and two in three within a week.

All scores are Amplitude's evaluations as relayed in OpenAI's post. The compared models are unnamed, so the results can't be reproduced from this material.

Human review stays in the loop

Wave outputs a recommendation with a proposed fix, a technical plan and acceptance criteria. Teams choose how much automation to allow. Per the story, most teams send bug fixes straight to coding agents, while new features and bigger bets get a human review.

Why it matters

The useful part for developers is the structure rather than the headline numbers. Each task is evaluated on quality, latency and cost, and the handoff to a coding agent carries acceptance criteria. A possible use of the same pattern is routing repetitive checks to a small model and keeping a stronger reasoning model for investigations.

The Data Science Agent result also shows why per-task evaluation helps: GPT-6.1 Sol scored slightly lower overall than the other model but cost less and scored higher on three criteria. For related OpenAI coverage, see the GPT-6 Cyber release and the Businesses in ChatGPT launch.