Claude Opus 5.5 is out: 40% cheaper to run, stronger on code

It matches Fable 5.1 on most tasks, beats GPT-6 Astra on several coding benchmarks, runs 30% faster, and writes less 'Claudish.' Sonnet and Haiku land in weeks.

Nowline SEP 22 11:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Fable-class coding, at Opus prices

    Opus 5.5 jumps to 66.4% on Terminal-Bench 4.0 (from 52.3%) and 57.8% on CursorBench 4.0 (from 46.6%), matching Claude Fable 5.1 on most tasks and topping GPT-6 Astra on several coding and agent benchmarks. Anthropic aims it at long, sprawling jobs like codebase-wide migrations.

  • 40% cheaper to run, and faster

    Input falls to $4/M and output to $20/M (from $5/$25), and cache reads drop 60% to $0.20/M — a real saving for agent loops that re-read context every turn. With ~30% faster generation and fewer tokens per task, Anthropic pegs typical workloads at 40% less than Opus 5.

  • Live today in Claude Code and the API

    The model ID is claude-opus-5-5, available now on the Claude Platform, AWS, Google Cloud and Azure, with a Fast mode ($8/$40 per M) in Claude Code for latency-sensitive work. Sonnet and Haiku versions are due in the coming weeks.

  • Tuned to sound less 'Claudish'

    Anthropic says the model front-loads key information, leans on less jargon and follows writing instructions more closely — a direct fix for the 'Claudish' critique. It also posts the lab's best automated behavioral-audit scores, with 85% fewer attempts to cross containment boundaries.

  • Elsewhere: 5B robotics models, real-time on a Jetson

    InstinctFlash, a new open-source serving runtime, claims real-time inference for 5B world-action (robotics) models on NVIDIA's Jetson Thor — putting frontier-scale embodied models on an edge board and a weekend project within reach.