AlphaEvolve Now Available to All Google Cloud Customers as a Full Product

Google Cloud has announced general availability of AlphaEvolve, a coding agent built on Gemini models that automatically optimizes algorithms through an evolutionary approach. The tool, which originated from Google DeepMind research, has moved onto the Gemini Enterprise Agent Platform after a year of private testing and is now available to every Google Cloud customer.

How the tool works

AlphaEvolve combines the creativity of Gemini language models with automated evaluators that check the correctness of generated solutions. Users supply a base program with specific code sections marked as open to optimization, along with a deterministic evaluation script that compiles, tests and scores successive variants. The system iteratively mutates the code, keeping the best ideas and discarding those that fail to improve the score.

Google describes the deployment process in four steps: define the problem and the base algorithm, set up an evaluation function that scores candidates for correctness and performance, run the optimization, and finally deploy the finished algorithm to production. The evaluation functions run on the client's side, on their own infrastructure, while AlphaEvolve generates candidates remotely via API.

Results from customers

Google Cloud published a list of companies that tested the tool in its early access program. BASF improved supply chain planning by 80 percent using a digital twin optimized by AlphaEvolve. FM Logistic shortened warehouse routes by 10.4 percent, saving 15,000 kilometers of travel. Schrödinger sped up molecule discovery fourfold, and PacBio reduced errors in genetic variant detection by 30 percent.

Inside Google itself, the tool was used to design the next generation of silicon for TPU chips, improve natural disaster risk prediction across 20 categories by 5 percent, and find quantum circuits with a tenfold lower error rate for the Willow quantum processor. Google also signed a partnership with the US Department of Energy under the Genesis Mission program and deployed AlphaEvolve on the Frontier supercomputer at Oak Ridge National Laboratory to generate GPU kernels.

Limitations of the tool

The tool only works where a measurable, fully automated function exists for evaluating the quality of a solution. Without such a metric, AlphaEvolve has no way to judge which code variant is better, making it unsuitable for tasks based on subjective judgment or ambiguous criteria. All the results cited by Google and its customers come from promotional materials and testimonials, without independent benchmark testing.

AI is moving from being a productivity-boosting assistant to a discovery engine that expands what we're able to achieve. By autonomously exploring complex computational spaces, tools like AlphaEvolve help researchers and engineers uncover breakthrough algorithms that complement human intuition - Pushmeet Kohli, Chief Scientist at Google Cloud and Vice President of Science at Google DeepMind

What this means for developers

For engineering teams, AlphaEvolve belongs to a different category of tool than popular coding assistants like Copilot or Cursor - it doesn't write code from a feature description, but instead explores the space of possible implementations for one tightly defined problem and returns the highest-scoring variant. Google has released documentation, a GitHub repository with Colab examples, and plugin-style integrations for tools such as Antigravity and Claude Code.

Access goes through the Gemini Enterprise Agent Platform console, with a free trial option. Google did not disclose pricing in its general-availability announcement, which makes it hard for now to judge whether the tool will pay off for smaller companies or remain the domain of large enterprise customers with the resources to build their own evaluation functions.

For Polish companies weighing adoption, the key factor will be having well-defined, measurable optimization problems, in logistics, production planning or IT infrastructure performance, where an automated evaluation function can actually be built. That sets AlphaEvolve apart from application code-generation tools, which today reach a much broader base of developers.