---
title: "Opus 5.5 is the best Opus since 4.5"
url: https://raskell.io/notes/opus-5-5-is-the-best-opus-since-4-5/
author: Raffael Schneider
date: 2026-10-24
description: "A month in, Claude Opus 5.5 feels like the next milestone after Opus 4.5: the same jump in what Claude Code can carry, now faster and cheaper per task."
tags: [ai-engineering, inference, economics]
---

# Opus 5.5 is the best Opus since 4.5

| | Opus 4.5 | Opus 5.5 |
|---|---|---|
| Released | 24 November 2025 | 22 September 2026 |
| Price per million tokens, input / output | $5 / $25 | $4 / $20 |
| Fast mode | not offered | $8 / $40 |
| Context window | 200K | 1M |
| Model ID | `claude-opus-4-5-20251101` | `claude-opus-5-5` |

Opus 4.5 was the moment Claude Code started working the way it was meant to. I had paid for Claude Code since May 2025 without being convinced; over the 2025 Christmas break, with Opus 4.5, it carried whole projects, and I built [Zentinel](https://zentinelproxy.io) with it. I wrote that up in [How I Work These Days](https://raskell.io/articles/how-i-work-these-days/) and [What Sixteen AI Agents Taught Me About Management](https://raskell.io/articles/what-sixteen-ai-agents-taught-me-about-management/), and it is the opening of my [Basel talk](https://raskell.io/talks/iat-13-basel/).

Opus 5.5 is the first release since then that felt like the same kind of step rather than an increment.

- **It holds the thread.** Long sessions that span a theme, a site, CI, and several data files stay coherent. The search across this site, the reading log, and the agent-readable files behind them were built in sessions with it.
- **Speed changes how I work with it.** Fast enough that I stay in the loop instead of switching away while it runs. Fast mode exists for when that matters more than the bill.
- **The cost moved the right way.** A cheaper rate card than 4.5, and Anthropic puts it at about 40 per cent cheaper to run than Opus 5 on typical work. The rate card is half the story; fewer wasted turns is the other half, the same argument I made in [The Economics of Inference](https://raskell.io/articles/the-economics-of-inference/).
- **Opus is the default again.** For most of my work I no longer pick a smaller model to save money, which was the point of 4.5 too.

What I do not know yet: whether it holds up on the hardest, longest-running tasks, where the gap to a larger model would show first. A month is not a benchmark.

Sources: [Introducing Claude Opus 4.5](https://www.anthropic.com/news/claude-opus-4-5), [Introducing Claude Opus 5.5](https://www.anthropic.com/claude-opus-5-5), [Opus 5.5 model overview](https://platform.claude.com/docs/en/models/opus-5-5/overview).