---
title: 'In the News: September 24, 2026 (Extra 3)'
description: 'Ruby on Rails published its own benchmark of 25 coding models on real feature tickets; two-day-old Claude Opus 5.5 ranks third behind two GPT-6 Astra configurations.'
canonical_url: 'https://darkfactory.dev/news/2026-09-24-extra-3'
markdown_url: 'https://darkfactory.dev/news/2026-09-24-extra-3.md'
collection: news
date_published: '2026-09-24T23:20:00-04:00'
date_modified: '2026-09-24T23:20:00-04:00'
---

# In the News: September 24, 2026 (Extra 3)


Ruby on Rails put its own framework to the test against 25 AI models on real feature work, not toy problems, and the newest arrival does not win.

## 1. Rails benchmarks 25 models on real feature tickets, and Opus 5.5 lands third

**[Rails and AI](https://rubyonrails.org/ai)** · Ruby on Rails core team · rubyonrails.org, accessed September 24, 2026

Rails' own "feature tickets" benchmark runs 20 tickets against Fizzy, 37signals' kanban app, written the way a product person would file them. Each model runs three times at the provider's default effort and again at maximum effort, capped at 90 minutes, 400 steps, and 60 dollars per run. GPT-6 Astra at maximum effort leads with 53.3 percent of runs passing both the app's test suite and the ticket's hidden checks, in a median 23 minutes 55 seconds. GPT-6 Astra at medium effort is second at 35.0 percent, and Claude Opus 5.5 at medium effort is third at 33.3 percent, taking 11 minutes 25 seconds for $2.78 a run. The methodology page and the raw runs sit in Rails' own open evaluation suite, linked from the benchmark. Rails did not independently disclose whether any outside party has rerun the suite, and this edition did not rerun it either.

The page also carries a quote from David Heinemeier Hansson: "Convention over configuration set the path for 20+ years of great training data for AI to use today." The framework's own numbers only partly back that claim. Even the top model still fails a majority of tickets.

**Why it matters:** For a team pairing a coding agent with a Rails codebase, this is the first public data placing Anthropic's two-day-old Opus 5.5 against the field on real feature work rather than general benchmarks, and third place is a reminder that a cheaper, faster model is not automatically the right one for a specific stack.