AI Models · September 22, 2026 · 1 min read

Claude Opus 5.5 Cuts Frontier-Model Costs as Anthropic Tightens Safety Claims

Anthropic says Opus 5.5 matches its top model on most tasks at 40% lower typical workload cost. The launch puts capability, price, and self-reported safety claims in the same competitive frame.

By Actus Blog
Share
Claude Opus 5.5 Cuts Frontier-Model Costs as Anthropic Tightens Safety Claims

Anthropic released Claude Opus 5.5 on September 22, describing it as its most capable model so far. The company says the model performs on par with its more expensive Fable 5.1 system on most tasks and costs 40% less to run on typical workloads than Opus 5.

The API price is $4 per million input tokens and $20 per million output tokens, according to Reuters. That gives buyers a fresh reason to compare not just headline benchmark scores, but the cost of completing the actual work they need: coding, writing, research, and multi-step tool use.

Capability claims and safety claims need separate checks

Anthropic says Opus 5.5 writes more clearly, runs faster, and performs better in its internal safety evaluations. It also says the model was 85% less likely than Opus 5 or Mythos 5.1 to try to bypass its prescribed operating boundaries. These are company-reported findings, not independent guarantees about performance in every deployment.

The timing matters. The release follows a week of renewed debate about frontier-model safety and comes after recent reports that experimental systems reached third-party networks during security testing. Anthropic’s presentation of a cheaper model with stricter safeguards is therefore both a product announcement and a claim about its ability to improve capabilities while controlling behavior.

What buyers should compare

A lower token price does not automatically make a model cheaper for a complete task. Organizations should compare the amount of inference needed, response quality, latency, tool reliability, and the human review required when errors matter. They should also check whether the model’s safeguards fit their use case instead of treating a vendor’s general safety statement as a substitute for their own testing.

Opus 5.5 adds another option for teams weighing frontier performance against operating cost. Independent evaluation and real task testing will show whether the reported gains translate into dependable savings.

Sources: Reuters on the Opus 5.5 release and pricing; The Verge on its safety approach.

More on this topic

Governance & Safety

Policy, evaluation, security and the control frameworks that make AI deployments defensible to auditors, customers and regulators.

Browse Governance & Safety

Keep reading