Opus 5.5 first impressions: strong coding, 20% cheaper output
Dan Shipper's early review finds Opus 5.5 strong at end-to-end coding and top on reading ease, but prone to doing too much and running past time limits, at $5/$20 per million tokens.
TL;DR: Opus 5.5 first impressions by @danshipper •Verdict: A strong daily-driver contender. It sometimes beats Fable 5.1 in Every's testing and is pulling some Codex converts back to Claude. •Coding: Excellent at end-to-end builds that match your taste. The author now picks it over Fable for big coding projects, and Kieran Klaassen swapped Fable 5.1 for it in daily product work ("new favorite model"). •Writing: Highest reading-ease score they've tested (68.42), so the output is easy to follow. It still buries the main point in intros and revisions, and trails Astra for everyday writing. •Pricing: $5/M input, $20/M output. Input matches Opus 5 ($5/$25), and output is 20% cheaper. Every estimates about 40% total savings. •Weakness: It has a "do the most" habit. Knowledge-work results were great, but at least one test ran past the 10-minute limit. •Author's split: Still roughly 80/20 Codex vs Claude, and still prefers Sol and Astra for editing. Claude's share of his tokens is up sharply week over week. •State of play: Codex remains the better harness for him, but Anthropic is gaining ground. It's repeating a pattern where a smaller model outperforms a bigger one, as with Sonnet 3.7.
This post quotes another post: x.com/i/web/status/2102435870309769352