AlphaEvolve
by Google
Gemini-powered evolutionary agent that discovers faster algorithms for your hardest optimization problems
AlphaEvolve is Google Cloud's evolutionary code-optimization agent. You supply a baseline algorithm and a deterministic scoring function, and Gemini-generated mutations are searched, scored and refined until a measurably better implementation emerges. It reached general availability in July 2026 and targets NP-hard optimization work in logistics, chip design, HPC kernels and quantitative modelling.
AlphaEvolve is Google's evolutionary coding agent, productised from Google DeepMind research and made generally available on the Gemini Enterprise Agent Platform on 9 July 2026 after a private preview that began in December 2025. It is emphatically not a general-purpose developer assistant — Google's own documentation states it is "not intended for basic code optimization like cleaning up code according to software engineering best practices" and targets algorithmic discovery, mathematical search and combinatorial optimization, typically NP-complete or NP-hard problems where many candidate implementations are functionally correct but only a subset meet the global performance criteria. The workflow has four steps: define a seed program with the code segments open to mutation plus problem context; define a deterministic evaluator that compiles, tests and scores candidates on scalar metrics; run the agentic harness, which queries the AlphaEvolve API for Gemini-authored mutations and feeds scores back so the search improves; then deploy the winning algorithm. Architecturally the evaluator runs client-side on the customer's own infrastructure, cloud or on-premises, while candidate generation happens remotely — so proprietary code and data need never leave the customer environment. Output is human-readable source engineers can review and maintain, not a black-box model. Optimization targets span runtime, throughput, latency, accuracy, memory, FLOPs, storage footprint, hardware utilisation and cost per unit of throughput. Google published unusually specific customer results: Kinaxis reported 22%+ better forecasting accuracy with a 90%+ runtime reduction, Klarna doubled ML training throughput after exploring roughly 6,000 candidates in three weeks under regulated-finance reproducibility constraints, JetBrains gained 15-20% on complex IDE algorithms, FM Logistic improved warehouse routing 10.4%, PacBio cut genomics variant-detection errors 30%, and internally it reduced Google Spanner write amplification by 20%. Access is by API, with Colab examples on GitHub and an AlphaEvolve Skill available inside IDEs including Antigravity and Claude Code.
Platform, HPC and quantitative teams that already own a measurable, automatable scoring function for an expensive algorithm and have stopped getting gains from hand-tuning it.
Replaces weeks of expert hand-optimisation with an automated search that returns readable, deployable code, with customers reporting 10-90% gains in runtime, accuracy or cost.
At a Glance
- Category
- Developer Tools
- Pricing
- Contact for pricing, Usage-based
- Target Market
- CTOs, Platform Engineering Leaders, Data Scientists, HPC Engineers, Quantitative Researchers
- Deployment
- Cloud-first, API-based, Hybrid
- Founded
- 1998
- Headquarters
- Mountain View, United States
- Team Size
- 500+
- Customers
- Named production users include BASF, Klarna, JetBrains, Kinaxis, FM Logistic, Coolblue, Infineon, PacBio, Schrodinger, WPP and Oak Ridge National Laboratory
Key Features
- ✓Evolutionary search harness
Gemini generates candidate mutations of your seed program and iterates on the highest-scoring survivors across generations.
- ✓Client-side evaluator
The scoring script runs on your own infrastructure, so proprietary code and data never leave your environment.
- ✓Human-readable output
Returns reviewable source code rather than a black-box model, so engineers can verify, test and maintain the result.
- ✓Multi-objective scoring
Optimizes simultaneously across runtime, accuracy, memory, FLOPs, storage footprint and cost per unit of throughput.
- ✓IDE Skill integration
An AlphaEvolve Skill runs inside IDEs including Antigravity and Claude Code for experimentation without leaving the editor.
- ✓Language-agnostic input
Accepts blocks of code in a given programming language rather than binding teams to a single runtime.
- ✓Google Cloud API delivery
Shipped as an API on the Gemini Enterprise Agent Platform with Colab examples and a public GitHub repository.
Capabilities
Use Cases
- •Supply-chain forecasting
Kinaxis improved forecast accuracy by more than 22% while cutting model runtime by over 90%.
- •ML training throughput
Klarna doubled pipeline throughput and improved model quality after evaluating roughly 6,000 candidates in three weeks.
- •Warehouse routing optimization
FM Logistic improved picking routes by 10.4%, saving more than 15,000 kilometres of staff travel.
- •Developer tooling performance
JetBrains achieved 15-20% performance improvements on complex algorithms behind IDE code completion.
- •Scientific and HPC computing
Schrodinger reported a 4x speedup in molecular dynamics inference, and Oak Ridge generates optimized GPU kernels on Frontier.
Ideal For
Best For
- ✓Combinatorial optimization such as warehouse routing, scheduling and supply-chain planning where a scoring function already exists
- ✓GPU and HPC kernel optimization, including complex mixed-precision kernels on exascale systems
- ✓ML training and inference pipeline throughput tuning where model quality can be measured objectively
- ✓Semiconductor design and lithography simulation where runtime is the binding constraint on larger chips
- ✓Forecasting and quantitative model improvement across supply chain, finance and marketing analytics
Not Ideal For
- ✗General application development or code cleanup — Google explicitly states it is not a general-purpose developer assistant and not meant for routine refactoring
- ✗Business logic with fuzzy success criteria: without a deterministic, automatable evaluator there is nothing for the search to optimize against, which practitioners identify as the boundary most production teams hit
- ✗Teams unwilling to invest in a rigorous scoring harness — commenters note poor metrics let the search exploit unmeasured dimensions and return fast but subtly incorrect code
- ✗Organisations that cannot adopt Google Cloud, since access is gated behind the Gemini Enterprise Agent Platform
Integrations
Deployment
Market & Ratings
Named production users include BASF, Klarna, JetBrains, Kinaxis, FM Logistic, Coolblue, Infineon, PacBio, Schrodinger, WPP and Oak Ridge National Laboratory
Market Analysis
Pros
- ✓Rare in agentic AI for shipping quantified, named customer outcomes rather than adjectives — 22% accuracy and 90% runtime gains at Kinaxis, doubled training throughput at Klarna, 10.4% routing improvement at FM Logistic
- ✓The client-side evaluator design means proprietary code and data stay in the customer's environment, removing the usual blocker for IP-sensitive optimization work
- ✓Produces reviewable code engineers can verify and maintain, so results survive audit and reproducibility requirements in regulated industries
- ✓Google has run it against its own critical infrastructure — TPU silicon layout, Spanner compaction, compiler storage footprint — which is stronger evidence than a customer logo wall
- ✓Available inside existing IDEs via an AlphaEvolve Skill, lowering the barrier to a first experiment
Cons
- ✗Only works where the problem has a measurable, automatable evaluation function — practitioners identify this as the wall most production teams hit, since most software is business logic with unclear success criteria
- ✗Building the scoring harness is real, unbudgeted engineering work, and poorly designed metrics let the search exploit unmeasured dimensions and produce fast but subtly incorrect code
- ✗Every published performance figure is vendor-supplied; none has been independently verified
- ✗No pricing, quota or billing information was disclosed at general availability, in either the announcement or the developer documentation
- ✗Closed-source with few implementation details published, which critics argue hinders independent verification and collective progress; OpenEvolve exists as an open alternative but has no published head-to-head comparison
- ✗An agent autonomously rewriting code in critical systems is itself an attack surface — researchers have flagged the risk of subtle backdoors and evaluator feedback-loop exploitation being reinforced across generations
- ✗Locked to Google Cloud's Gemini Enterprise Agent Platform, which rules it out for teams standardised elsewhere
Pricing
Gemini Enterprise Agent Platform
Contact for pricing
- ✓AlphaEvolve API access
- ✓Client-side evaluator runner
- ✓Colab examples and public GitHub repo
- ✓AlphaEvolve Skill for Antigravity and Claude Code
Google published no list price for AlphaEvolve at general availability, and the developer documentation states neither quotas nor billing rates; it is delivered through Gemini Enterprise Agent Platform entitlements, so cost lands inside an existing Google Cloud contract and is negotiated there. Budget separately for the evaluation side of the loop, which is where real spend accumulates: the scoring harness runs on your own compute, and published customer runs span 200 to roughly 6,000 candidate evaluations, meaning wall-clock time and infrastructure cost scale with how expensive a single evaluation is.
Security & Compliance
Connect
Sources
This page was written from 6 sources, 5 on domains other than docs.cloud.google.com.
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Greptile
AI code review agent that indexes the whole codebase to catch cross-file bugs before merge
GitLens
The Git workbench where human and AI-agent changes get reviewed, composed and merged in one view
OpenHands
Open-source, self-hosted control center for running coding agents on real engineering work
Confident AI
Hosted LLM evaluation, observability and red teaming built on the open-source DeepEval framework