AlphaEvolve went GA on Google Cloud last Wednesday, out of the private preview that had been running quietly since December 2025. It ships as part of the Gemini Enterprise Agent Platform (still the fresh name for what was Vertex AI three weeks ago), and Google published a customer case-study list that is, for once, actually loaded with numbers rather than adjectives.
The numbers: Klarna doubled its AI throughput while improving model quality, and it did that by letting AlphaEvolve explore roughly 6,000 candidate programs over three weeks. BASF improved its existing planning and forecasting models by more than 80%. Kinaxis got a 22% lift in forecasting accuracy and a 90%-plus runtime reduction on benchmark workloads. JetBrains squeezed 15 to 20% more performance out of its IDE. Schrödinger got a 4x speedup on molecular discovery. FM Logistic cut warehouse-routing distance by 10.4% and saved staff about 15,000 kilometers of walking. Pebble reduced GPU performance-modeling error by 56%. Pricing is not on the blog post, which is Google’s way of saying it is expensive and negotiated.
The pitch is that AlphaEvolve is not a coding assistant. It is, in Google’s language, a discovery agent, and Pushmeet Kohli’s launch quote frames it as “moving beyond acting as a productivity assistant that accelerates how we work to a discovery engine that expands what we can achieve.” Cut through the org-chart prose and the mechanism is straightforward: you provide a baseline algorithm and a scoring function, AlphaEvolve mutates candidate programs against the scoring function in an evolutionary loop, and it hands you back the best version it found. It compiles, runs, and evaluates thousands of variants in a closed loop. Nothing about it is speculative. It is genetic algorithms plus a Gemini-scale mutation operator.
The important sentence in the whole announcement is the one about the four-step deployment process. Step one is “define the seed algorithm.” Step two is “measure by establishing a scoring function to objectively score candidate programs.” That step-two clause is doing an enormous amount of work. It means the customer has to write, and defend, a deterministic client-side evaluator that tells AlphaEvolve which candidate is better. If you cannot write that evaluator, AlphaEvolve cannot help you. If your evaluator is wrong or gameable, AlphaEvolve will happily produce a program that maximizes the wrong number.
That is the actual product shape, and it explains why the customer list is what it is. Klarna, BASF, Kinaxis, FM Logistic, Schrödinger, Pebble: these are organizations that already know exactly what they mean by “better” in their domain and can encode it. Warehouse routing has an unambiguous fitness function (total kilometers walked, subject to constraints). Molecular docking has a fitness function. Forecasting accuracy has a fitness function. IDE cold-start time has a fitness function. Every one of the reported wins is a shop where the scoring rule was already crisp before AlphaEvolve arrived. The model did the searching. The customer defined the destination.
Which is the interesting long-term implication for the coding-agent category. The GitHub Copilot and Cursor line of tools optimizes for the case where the human knows what “better” looks like case by case and is comfortable eyeballing suggestions. AlphaEvolve optimizes for the case where you can precommit, in writing, to what better means, and are willing to let a machine spend a week rearranging your code inside that constraint. Those are different products for different problems, and the second one is going to be a much smaller market with much better unit economics. Google is now selling both.
The unsaid quiet part is that AlphaEvolve is also a very effective way to burn Gemini tokens against a real business outcome, which is exactly the shape Google wants its enterprise customers to develop a taste for right now. Every 6,000-candidate exploration is a lot of Gemini calls billed to somebody’s cost center. Klarna got its throughput doubled. Google got a customer who now has an internal number for what a doubled-throughput Gemini spend is worth. Both sides win. The Gemini Enterprise Agent Platform pitch depends on there being more customers who understand what a Gemini run is worth in dollars, and this is how you manufacture that understanding at scale.