Anthropic Launches Opus 5.5 With Lower Costs
Anthropic has launched Claude Opus 5.5, the first model in its new Claude 5.5 family. The company says it delivers performance comparable to Claude Fable 5.1 across most work. Moreover, Anthropic says typical workloads cost 40% less than Opus 5.
The model arrived on September 22, 2026. Meanwhile, Anthropic positioned Opus 5.5 for agentic coding, computer use, research, and professional knowledge work. The release also introduces stronger safeguards for long-running autonomous tasks.
Opus 5.5 Targets Higher Efficiency
Opus 5.5 costs $4 per million input tokens and $20 per million output tokens. Those rates are 20% lower than Opus 5. Additionally, cache reads fall from $0.50 to $0.20 per million tokens. Cache writes also decline from $6.25 to $5.
However, Anthropic’s larger cost reduction comes from efficiency. The company says Opus 5.5 uses fewer tokens to complete tasks. Consequently, overall running costs can fall by about 40% compared with Opus 5.
The model also offers a faster mode through Claude Code and the Claude Platform. Fast mode can reach up to 2.5 times the standard speed. However, it costs $8 per million input tokens and $40 per million output tokens.
Anthropic’s benchmarks show gains across several agentic coding tasks. For example, Opus 5.5 scored 66.4% on Terminal-Bench 4.0. It also reached 54.4% on FrontierCode v1.1 and 57.8% on CursorBench 4.0.
On GDPval-AA v2.1, Opus 5.5 recorded 1,846 Elo. That result exceeded Fable 5.1’s 1,735 and Opus 5’s 1,708. However, Anthropic notes that benchmark differences can become less reliable at these capability levels.
Coding and Knowledge Work Lead Use Cases
Opus 5.5 is designed for large and complex coding tasks. According to Anthropic, an early tester completed a 680,000-line code migration in less than one day. Another tester audited and fixed a 200,000-line codebase in under three hours.
Furthermore, Anthropic tested Opus 5.5 on a translation of HAProxy from C to Rust. Both Opus 5.5 and Fable 5.1 produced rewrites that passed nearly all of HAProxy’s regression tests. However, Opus 5.5 completed the task faster and at 51% lower cost in the company’s test.
The model also targets broader professional workflows. Anthropic tested it on research reports, financial analysis, business automation, and large-scale data collection. In one internal test, 16 of 18 Opus 5.5 reports met the company’s quality threshold.
At the same time, the model includes tools aimed at safer autonomous coding. Anthropic says a classifier screens every action before it is executed. The company also provides an auditable sandbox and automated code review for security vulnerabilities.
Stronger Safeguards Accompany New Model
Safety is a major part of the Opus 5.5 release. Anthropic says the model underwent external testing from Frontier Design and METR before launch. Furthermore, the company expanded its alignment testing to cover longer tasks and scenarios based on real incidents.
According to Anthropic, Opus 5.5 was 85% less likely than Opus 5 or Mythos 5.1 to attempt to circumvent containment boundaries in one internal evaluation. Anthropic says every such attempt was low severity and self-reported.
The company has also added safeguards for sensitive workloads. Certain cybersecurity requests can be redirected to Opus 4.8, while flagged biology requests can be routed to Opus 5.
The launch also follows Anthropic’s recent release of Fable 5.1 and Mythos 5.1. Those models arrived earlier in September as the company’s higher-end systems for coding, knowledge work, and research.
Anthropic says Claude Sonnet 5.5 and Claude Haiku 5.5 will follow in the coming weeks. Therefore, Opus 5.5 marks the start of a broader model-family update rather than an isolated release.
For businesses, the immediate change is the combination of lower token pricing and higher reported efficiency. Meanwhile, developers gain a model designed for longer coding and knowledge-work tasks. The practical impact will depend on how these performance and cost gains translate into production workloads.
