Anthropic released Claude Opus 5.5 on Tuesday, the first model in what the company is calling its Claude 5.5 family. Anthropic says the new model performs at the level of Claude Fable 5.1, its priciest flagship, on most tasks.

The twist is the model costs 20% less to run than Opus 5, the model it replaces, and it’s 60% cheaper than Fable 5.1, the company’s state-of-the-art offer.

The model, it appears, is very capable. It’s the most advanced model both in version (5.5 vs Fable’s 5.1) and family (Opus is the biggest model publicly available, followed by Sonnet with Haiku being the smallest one).

Anthropic admits the gap between Fable and Opus is narrower than its benchmark scores suggest, but being publicly available on plan, and being cheaper per token, makes this the obvious choice for most Claude users.

Opus 5 launched in July underpricing and outscoring Fable 5 on most benchmarks, and Fable 5.1 arrived in early September and reversed that, beating Opus 5 on every benchmark Anthropic published.

Opus 5.5 closes the gap again, this time on price rather than raw scores. Lovable, a developer platform that ran Opus 5 through its own coding tests back in July, said in a statement shared by Anthropic that the earlier model was “steadier, with far less variance run to run” than what came before it—the same efficiency pitch Anthropic is making again now.

Opus 5.5 is also the first model Anthropic has shipped since CEO Dario Amodei published an essay calling on the industry to slow down, arguing that AI capability gains are outrunning the safety testing meant to keep them in check. “We must slow the pace at which we improve the capabilities of AI models,” Amodei wrote earlier this month.

Anthropic measures most of this progress through agentic coding—work carried out by an AI agent, a model that takes multi-step actions on its own instead of just answering a single prompt.

On Terminal-Bench 4.0, which tests whether an agent can finish complex professional tasks inside a command-line interface and scores the result as a percentage of tasks completed, Opus 5.5 hit 66.4%, ahead of Fable 5.1’s 55.8% and GPT-6 Astra’s 57.9%. FrontierCode, which checks whether an agent’s code changes would actually get merged in a real engineering pipeline using that same pass-rate scoring, put Opus 5.5 at 54.4% against Fable 5.1’s 50.3%.

On GDPval-AA v2.1, which grades real-world professional work across 44 occupations using Elo—the same chess-style ranking system used to measure relative skill rather than a flat percentage—Opus 5.5 scored 1846 against Fable 5.1’s 1735 and Opus 5’s 1708.

Opus 5.5 doesn’t lead everywhere, though. On AutomationBench, a Zapier-built benchmark that scores whether an agent can carry a full business workflow start to finish without human help, GPT-6 Astra’s 41.4% edges out Opus 5.5’s 40.0%.

On Terminal-Bench-Science 0.1, which tests agentic scientific research carried out inside a command-line environment using that same pass-rate method, Astra’s 64.6% beats Opus 5.5’s 58.7% by a wider margin.

The bigger selling point is cost. Input tokens—the chunks of text a model reads and writes, priced per million—now run $4 for Opus 5.5 versus $5 for Opus 5, and output tokens drop to $20 from $25. Cache reads, which mean reusing context a model has already processed instead of reprocessing it from scratch and driving most of the cost on long agentic coding sessions, fall 60% to $0.20 per million tokens.

Because Opus 5.5 matches Anthropic’s Mythos-class models—the company’s most restricted tier, reserved for vetted cybersecurity and biology researchers—on those two capabilities, it launches with the same safeguards as Fable 5.1.

Most cybersecurity tasks get automatically rerouted to the older Opus 4.8, which is something many devs and users have previously complained about..

Opus 5.5 is live now across Anthropic’s platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure. Sonnet 5.5 and Haiku 5.5 will follow in the coming weeks, Anthropic says, carrying the same price cuts down the rest of the lineup.