OpenAI published a customer story about Amplitude on November 12, 2026, and the page is more specific than the genre allows: it puts numbers on model cost per task. Amplitude says Wave, its proactive product agent, ships product improvements in an estimated one to two days instead of roughly a month, and that the agent runs about 80% on OpenAI models. The claim that matters for anyone paying for agent tokens is not the 160% engagement lift at the top. It is that Amplitude sends different jobs to different models on purpose, and publishes what that split costs.
What was actually released
Nothing was released. The page is OpenAI’s account of one customer, filed under North America, technology, mid-market, with Codex as the product in use.
The workflow explains the numbers. A team gives Wave a product area plus the goals and metrics it wants to move. Wave pulls Amplitude data, session replays, customer feedback and code into one pass, ranks the opportunities, drafts a solution with a technical plan and acceptance criteria, then sends that plan to a coding agent, Codex included.
What the page actually commits to
- Amplitude estimates one to two days to ship an improvement with Wave, against roughly a month before. The word is estimates: no sample, no start date, no count of improvements.
- Three OpenAI models do three different jobs, and a fourth product, Codex, does the implementation. That routing, not the model choice, is the operating idea.
- The human keeps the gate on anything new. The page says most teams send bug fixes straight to coding agents, while new features and bigger bets go through review.
The split by task is explicit. GPT-6.1 Sol runs the opportunity manager and the deep investigations into code and data. GPT-6 Luna takes the frequent, cheap work: scanning code, running small evaluation checks. GPT-Image-2.5 Flare generates product mockups from screenshots of session replays. The page says the team evaluates quality, latency and cost for each task, then assigns the model.
What the claim rests on
Every number on the page is an evaluation Amplitude ran on its own product, published by its supplier. That does not make the numbers wrong, and the page is more candid than most customer stories: it publishes one evaluation where its chosen model loses on quality. It does mean nothing here is an independent measurement, and no artifact behind the scores is available.
| Evaluation | GPT-6.1 Sol | Comparison | Unit |
|---|---|---|---|
| Opportunity Investigator, overall | 4.06 | 3.62, 3.52 and 2.94 for three other models | out of 5 |
| Opportunity Investigator, median time | 54 minutes | not published for the others | per investigation |
| Data Science Agent, quality | 67.5 | 69.3 for another leading model | out of 100 |
| Data Science Agent, mean model cost | $1.28 | $2.63 | per dispatch |
| Solution wireframes, GPT-Image-2.5 Flare | 80.3 | 63.9 for another image model | out of 100 |
The 160% is narrower than it reads. On Amplitude’s own homepage, suggestion cards meant to start a conversation with an agent were mostly ignored. In a sample of 100 session replays, 80 users typed their own question and 12 picked a suggestion. Wave proposed personalizing the cards from recent charts, dashboards or the page a user came from, kept free typing in place, and the team A/B tested the change. Messages sent through those cards rose 160% against control after one day. Traffic, confidence interval and test length are not on the page.
The same pattern holds for the launch timeline. Among customer-launched experiments, flags and guides, one in four went live the same day Wave surfaced the opportunity, half inside four days, two in three inside a week. The page itself footnotes that curve as interpolated from supplied milestones and illustrative, not individual observations.
What is not published
No token count per dispatch, so a model cost cannot be rebuilt. No total spend, no price for Wave, no seat count, and no named customer outside Amplitude itself. No independent audit of the A/B test or of any of the five evaluation scores. And no definition of the baseline behind roughly a month.
It is a great example of how non-groundbreaking things that just get ignored can increase the adoption of features in unexpected ways.
That line comes from Eric Carlson, senior principal AI engineer at Amplitude, and it is the most useful sentence on the page. The 160% did not come from a better model. It came from noticing that users were typing instead of clicking.
On this page
What it changes for product teams paying agent tokens
Two budget decisions get clearer.
The first is per-task routing. Route by job, not by generation: Sol at $1.28 per dispatch against $2.63 for the model it beat on cost is a 51% cut on the same evaluation, and it frees the expensive model for the investigation step.
| Model | Job in Wave | Figure attached |
|---|---|---|
| GPT-6.1 Sol | opportunity manager, deep code and data investigations | $1.34 model cost in the Opportunity Investigator evaluation |
| GPT-6 Luna | code scans, small evaluation checks | none |
| GPT-Image-2.5 Flare | mockups from session replay screenshots | 80.3 out of 100 on wireframes |
| Codex | implementation once a plan is accepted | none |
The second is what a dispatch is worth. Sol’s rate on our record is $2 per million input tokens and $10 per million output, from the supplier’s own Terra announcement, where Terra at $1 and $5 is priced as half of Sol. A dispatch that costs $1.28 at those rates is not a chat message, it is a run that reads code and data. Anyone budgeting an agent on message volume will be off by an order of magnitude.
What it does not change is the gate. The page keeps review on features and releases bug fixes to agents, which is a policy most product teams can copy without buying anything. The useful number here is not 1-2 days. It is the split between what ships unattended and what waits for a person.
The Neural Minute take
Read the page as a comparison of models and you learn the wrong thing. Read it as a comparison of tasks and it says something a buyer can act on: the supplier’s customer publishes an evaluation where the flagship scores 67.5 against 69.3 for a rival, and still routes to the flagship there, on cost per dispatch. That argues for keeping two models live in your own evaluation instead of migrating everything when one model improves.
The other thing worth saying is that Wave did not find the lift. Amplitude did, by reading 100 session replays and noticing that 80 users typed and 12 clicked. That work needs product analytics, not a particular model, and it is reproducible on any stack. What the models did was shrink the distance between that observation and a tested change, from a month of tickets to a couple of days. We read the same pattern in the GPT-6.1 Terra price cut: the catalog is segmented by effort, so the buyer’s decision is a setting rather than a model.
So instrument the product so that 100 replays produce a finding, keep a human gate on anything new, and price the agent per dispatch rather than per message. Skip the idea that one model carries an agent end to end: the supplier’s own customer story shows the opposite.
Sources
- How Amplitude ships product improvements in 1-2 days with OpenAI, OpenAI, published November 12, 2026: https://openai.com/business/customer-stories/amplitude
- Every figure above is read on that page on November 13, 2026 as Amplitude’s and OpenAI’s own statement, not as an independent measurement
- Our reading of the GPT-6.1 Terra price cut and the rate we hold for GPT-6.1 Sol, October 27, 2026: https://www.neuralminute.com/news/gpt-6-1-terra-openai-price-cut/