Claude Opus 5.5

Claude Opus 5.5 replaces Opus 5 as Anthropic’s Opus-tier model, released on 22 September 2026, two months after Opus 5. Anthropic’s pitch is that it performs at the level of Claude Fable 5.1 on most work while costing about 40% less to run than Opus 5. Sonnet 5.5 and Haiku 5.5 are announced “in the coming weeks.”

What changed

The per-token price came down: $4 per million input tokens and $20 per million output, against $5 and $25 for Opus 5. Cache reads dropped from $0.50 to $0.20 per million, which matters on long agent sessions that re-read the same context. Anthropic puts the total saving on typical workloads at about 40%, so the rest comes from the model getting through the same work with fewer tokens. GitHub reported the same in its early Copilot testing: comparable results to Opus 5 in fewer steps and tokens. Output also comes more than 30% faster, by Anthropic’s measure.

It also talks differently. Answers lead with the important part, use less jargon, and are noticeably shorter.

The context window stays at 1M tokens and output at 128K. Fast mode, for when you need speed more than savings, runs at $8 and $40 per million and is only available on the Claude API.

What to know before you migrate

Four changes will break code written against Opus 5. Thinking can no longer be turned off at any effort level, so effort is the only setting you have. The default effort also dropped from high to medium. Forced tool use returns an error. Thinking blocks are tied to the model and the conversation they came from. And computer use needs Anthropic’s newer computer toolset. If you work through the Claude apps or Claude Code rather than the API, none of this touches you.

Where it sits in the lineup

Right now this is the model I’d use for almost everything. Anthropic’s advice has been to start on Opus and move up to Fable only when your own evaluations fall short, and Opus 5.5 makes that step rare. It matches Fable on most of Anthropic’s benchmarks and beats it on some: on Artificial Analysis’s GDPval-AA, a test of real-world work across 44 occupations, it scores 1846 Elo against Fable 5.1’s 1735. Fable costs two and a half times as much per token. What’s left for Fable is a little extra edge on very big, long-running tasks. That’s a snapshot: the next Fable release will probably move the line again.

Further down the range, I no longer pick Sonnet 5 by hand at all. It still does plenty of work as a subagent, when Opus hands it a focused job, and at half Opus’s per-token price that’s where it belongs. Haiku is down to one job for me: searching large document collections as a subagent.

On a subscription, the per-token price is the wrong thing to look at. What matters there is how fast a model eats your usage allowance, and this one eats it more slowly than Opus 5 did. Anthropic also raised the five-hour usage limits on Pro, Max, Team and seat-based Enterprise plans with this release, and it’s now the default model in the Claude app, Claude Code and Cowork.

Prompt it for outcomes

The advice from Opus 5 still holds, and more so here. These models do their best work when you describe what you want to achieve and how you’d know it worked, and leave the route to them. A prompt that spells out every step gets you a model following steps. A prompt with a clear goal and a test for it gets you the new approaches this one is good at finding.

The Verdict

Best for: Almost everything: planning, deep reasoning, writing, research, and complex agentic coding. Fable only adds an edge on very big, long-running tasks.