This is the third part of the series about Ruby's future, inspired by Irina Nazarova's keynote "Startups on Rails" at RubyConf Thailand 2026. Earlier we discussed myths and fears among engineers (Part 1) and real AI startups (Part 2). Today we move to technical facts and benchmarks.

The old myth that "Ruby is slow" dates back over a decade. Let's see where things stand in 2026, with Ruby 3, the YJIT compiler, and LLM-based code generation.

Fact 1: Ruby Is ~2.5x Faster Than Python

Ruby 3 introduced massive speedups over Ruby 2 (better memory management, GC, M:N threads), and YJIT by Shopify adds another 15-20% performance boost with a single flag.

Latest benchmark results from bddicken/languages (run on an M3 MacBook Pro with 16GB RAM):

  • 1 Billion Nested Loops: Ruby (3.3.5) 28.80s vs Python (3.9.6) 74.42s — Ruby is 2.58x faster
  • Naïve Fibonacci (N=40): Ruby (3.3.5) 12.17s vs Python (3.9.6) 29.00s — Ruby is 2.38x faster

Across general computation, standard Ruby runs ~2.4x–2.6x faster than standard CPython.

Comparing micro-frameworks (TechEmpower Round 23, Fortunes test):

       Micro-Frameworks (Requests per second)

  Express.js (Node.js)  │ ██████████████████████████ 172.5K req/s
  Roda (Ruby + Iodine)  │ █████████████████████████  165.7K req/s
  FastAPI (Python)      │ ████████████████           109.1K req/s
  FastAPI + ORM (Python)│ █████                      37.8K req/s

Roda on Ruby runs at 96% of Express.js speed (165.7K vs 172.5K req/s) and easily leaves Python's FastAPI far behind.

Full-stack frameworks (same source):

       Full-Stack Frameworks (Requests per second)

  Laravel Workerman     │ ██████████████████████████  49.5K req/s
  Rails Iodine          │ ████████████████████████    47.9K req/s
  Rails Puma (default)  │ ██████████████████████      42.5K req/s
  Rails Falcon          │ █████████████████████       40.6K req/s
  Django (Python)       │ ████████████████            32.6K req/s
  Laravel FPM (default) │ ████████                    16.4K req/s

Even default Rails (Puma at 42.5K req/s) beats optimized Django (32.6K req/s). And standard Rails outperforms default PHP/Laravel FPM (16.4K req/s) by 2.6x.

Fact 2: Token Savings in Code Generation (Token Efficiency)

In the era of AI agents (Cursor, GitHub Copilot, Claude Code), a language's conciseness directly affects context window usage, cost, and agent speed.

LLMs process code via tokens. The more tokens required to express logic, the faster you exhaust the context window and the higher the risk of model drift.

According to Martin Alderson's study on language token efficiency (using the GPT-4 tokenizer across RosettaCode tasks):

  • Ruby averages 116 tokens per task — ranking #3 overall among major programming languages (beating Python, JavaScript, Java, Go, Rust, and C#).
  • Python: 128 tokens (~10% more tokens than Ruby).
  • JavaScript: 181 tokens (~56% more tokens than Ruby).
  • Java (187 tokens), Rust (198 tokens), Go (207 tokens), C# (211 tokens): Require 60% to 80% more tokens than Ruby for the exact same task.
  • C++ (264 tokens) & C (285 tokens): Require over 2.4x more tokens than Ruby.
┌─────────────────────────────────────────────────────────────┐
│  Average Tokens per Task (GPT-4 Tokenizer, Lower = Better)  │
├─────────────────────────────────────────────────────────────┤
│ Ruby        █████████ 116 tokens                            │
│ Python      ██████████ 128 tokens                           │
│ JavaScript  ██████████████ 181 tokens                       │
│ Java        ███████████████ 187 tokens                      │
│ Rust        ████████████████ 198 tokens                     │
│ Go          █████████████████ 207 tokens                    │
│ C#          █████████████████ 211 tokens                    │
└─────────────────────────────────────────────────────────────┘

Alderson followed this up by analyzing web frameworks, demonstrating that framework choice matters even more than language choice for AI agents. In Rails, standard conventions remove boilerplate, keeping prompt context clean and focused on business logic.

Fact 3: Convention over Configuration Is a Natural Fit for AI Agents

How do AI agents work when writing code for other stacks? You end up creating massive AGENTS.md or CLAUDE.md files, explaining to the model where controllers live, how to name routes, and how to structure the database. There are now over 60,000 repos with AGENTS.md files telling AI about project structure.

In Rails, the principle of Convention over Configuration means:

  1. AI models already know Rails conventions because they were trained on 20 years of open source repositories following the same patterns.
  2. The architecture is predictable: the model knows exactly that User lives in app/models/user.rb.
  3. No CLAUDE.md needed to explain where models, controllers, and tests go. The framework is the prompt. The agent focuses only on business logic.

Architectural Pragmatism: Dropping the Purism

Nazarova stresses: to build fast systems in 2026, the community needs to let go of purism.

  ┌──────────────────────────────────────────────────────────┐
  │              Modern Rails Application Stack              │
  ├───────────────────────┬──────────────────────────────────┤
  │ Business logic        │ Ruby on Rails (maximum velocity) │
  ├───────────────────────┼──────────────────────────────────┤
  │ Heavy Async / Sockets │ AnyCable / Go / Rust             │
  ├───────────────────────┼──────────────────────────────────┤
  │ ML models             │ Python (via API / gRPC)          │
  ├───────────────────────┼──────────────────────────────────┤
  │ Frontend              │ React, Vue, Svelte via Inertia   │
  └───────────────────────┴──────────────────────────────────┘

Databases are written in C++, heavy build tools in Go/Rust, and ML model logic runs on Python. The infrastructure tools should know nothing about your business logic — they just provide performance. Rails stays the brain of the application, where the business logic lives, easy for both humans and AI to read and write.

➡️ Read Part 4: What Ruby Is Missing for AI and How the Community Can Take Back the Initiative

Have thoughts or feedback on this dispatch? GET IN TOUCH →