Summary
- On Tuesday, Anthropic unveiled Claude Opus 5.5, priced at $4 and $20 per million tokens, making it 40% cheaper than its predecessor, Opus 5, while matching Fable 5.1 in most tasks.
- According to Anthropic’s data, Opus 5.5 excels over Fable 5.1 and Opus 5 in coding and knowledge-related evaluations, although GPT-6 Astra outperforms it in AutomationBench and Terminal-Bench-Science 0.1 tests.
- This release marks the first model since CEO Dario Amodei urged the industry to moderate AI advancements, and it includes the same cybersecurity and biological safeguards as Fable 5.1.
On Tuesday, Anthropic introduced Claude Opus 5.5, the initial release in what the firm refers to as its Claude 5.5 family. The company asserts that this model performs comparably to its top-tier Claude Fable 5.1 in most tasks.
Notably, the new model operates at a cost that is 20% lower than Opus 5, the model it replaces, and is 60% less expensive than Fable 5.1, the company's premium offering.
Myriad: Which company IPOs next? Click to make your prediction.Opus 5.5 appears to be highly capable, marking a significant advancement over its predecessor in both version (5.5 compared to Fable’s 5.1) and family (Opus being the largest publicly available model, with Sonnet and Haiku as smaller options).
Anthropic acknowledges that the performance difference between Fable and Opus is narrower than what benchmark scores might imply, but the public availability and lower cost per token make it the preferable option for many users of Claude.
Launched in July, Opus 5 had initially outperformed Fable 5 across most benchmarks, but the arrival of Fable 5.1 in early September reversed that trend, with Fable 5.1 surpassing Opus 5 on all published benchmarks.
With Opus 5.5, Anthropic has again narrowed the gap, this time focusing on pricing rather than performance metrics. Lovable, a developer platform that previously tested Opus 5, noted that the earlier model was "steadier, with far less variance run to run," a claim Anthropic is reiterating with this new release.
Opus 5.5 is also the first model launched since CEO Dario Amodei published an essay calling for a slowdown in AI capability increases, asserting that the pace of advancements is outstripping necessary safety evaluations. "We must slow the pace at which we improve the capabilities of AI models," Amodei stated earlier this month.
Anthropic assesses most of its progress through the lens of agentic coding, which involves an AI agent performing multi-step tasks autonomously rather than simply responding to single prompts.
In the Terminal-Bench 4.0 evaluation, which measures an agent's ability to complete complex professional tasks via a command-line interface, Opus 5.5 achieved a score of 66.4%, outperforming Fable 5.1's 55.8% and GPT-6 Astra's 57.9%. In the FrontierCode assessment, which evaluates whether an agent's code modifications would be integrated into a real engineering workflow, Opus 5.5 scored 54.4%, compared to Fable 5.1's 50.3%.
However, Opus 5.5 does not excel in all areas. In the AutomationBench, a benchmark designed to assess an agent's ability to independently complete an entire business workflow, GPT-6 Astra scored 41.4%, slightly ahead of Opus 5.5's 40.0%. Meanwhile, in the Terminal-Bench-Science 0.1 test, which evaluates agentic scientific research in a command-line context, Astra's 64.6% score surpassed Opus 5.5's 58.7% by a notable margin.
The key advantage of Opus 5.5 lies in its affordability. The cost for input tokens—representing the text the model processes—has been reduced to $4 per million for Opus 5.5, down from $5 for Opus 5, while output tokens decreased from $25 to $20. Additionally, costs for cache reads, which enable models to reuse previously processed context and significantly reduce expenses during long coding sessions, have dropped by 60% to $0.20 per million tokens.
Opus 5.5 is designed to match Anthropic's Mythos-class models, which are reserved for vetted cybersecurity and biological researchers, and it launches with the same safeguards as Fable 5.1.
Most cybersecurity tasks will still default to the earlier Opus 4.8 model, a decision that has drawn complaints from many developers and users.
Currently, Opus 5.5 is available across Anthropic's platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure. Anthropic plans to release Sonnet 5.5 and Haiku 5.5 in the upcoming weeks, which will also feature similar price reductions throughout the product line.