Anthropic Launches Claude Opus 5 on AWS, First Fifth-Generation Opus Model

Anthropic's Claude Opus 5 is now generally available on Amazon Bedrock and Claude Platform on AWS, the companies announced. It is the first model in Anthropic's fifth-generation Opus line, succeeding Claude Opus 4.8.

AWS and Anthropic did not disclose a training cutoff date, context window size, or per-token pricing in the announcement. No standardized benchmark scores (MMLU, HumanEval, SWE-bench, etc.) were published alongside the release.

What Anthropic claims changed

According to Anthropic, Claude Opus 5 delivers improvements across four core areas: agentic coding, knowledge work, visual understanding, and long-running autonomous tasks. The company claims the model can navigate codebases like an experienced engineer, write production-quality code, and adapt its strategy mid-task. Anthropic also claims Opus 5 can run as a dependable agent for hours or overnight, recovering from errors and working around obstacles without human intervention.

For knowledge work, Anthropic claims the largest capability gains show up in document-heavy enterprise tasks — deeper reasoning over long documents and higher accuracy on complex analysis. The company also claims the model improves on Opus 4.8's cybersecurity capabilities "across the board," from coding to security-specific tasks.

Notably, Anthropic claims Claude Opus 5 matches the intelligence of its Claude Fable 5 model in many domains, but at what it calls "Opus-tier pricing" — implying a cost structure similar to prior Opus releases rather than a premium tier. No dollar figures for input or output token pricing were provided in the announcement; pricing not yet disclosed.

Anthropic also states that in unspecified "higher-risk areas," Opus 5 may automatically fall back to Claude Opus 4.8, with users notified when this occurs. Developers can configure fallback behavior through the API.

Enterprise and deployment details

On Amazon Bedrock, Claude Opus 5 ships with zero data retention (ZDR) enabled by default, running on what AWS describes as its "next-generation inference engine." AWS says this allows customers to keep workloads within existing AWS environments, maintain regional data residency, and scale inference with no operator access to underlying data.

The model is also available through Claude Platform on AWS, where zero data retention is available on request. This route gives customers access to Anthropic's native API and console feature set while billing and authentication route through AWS.

The model supports mid-conversation tool addition and removal via tool_addition and tool_removal content blocks, avoiding the need to resend a full tools array on every turn — a detail aimed specifically at long-running agentic workflows.

At launch, Claude Opus 5 is available on Amazon Bedrock in US East (N. Virginia), Asia Pacific (Melbourne), Europe (Ireland), and Europe (Stockholm), with additional regions listed in AWS documentation. Claude Platform on AWS supports the model across North America, South America, Europe, and Asia Pacific. The model ID for API calls is global.anthropic.claude-opus-5.

What this means

This release is notable less for new capability disclosures — none of the usual benchmark comparisons (SWE-bench, GPQA, MMLU) accompanied the announcement — and more for distribution strategy. Anthropic continues to deepen its dependency on AWS as a deployment surface, pairing model releases with enterprise-friendly guarantees like default zero data retention and regional residency, which matter more to regulated industries (financial services, healthcare) than raw benchmark leadership.

The claim that Opus 5 matches "Fable 5" intelligence at Opus-tier pricing suggests Anthropic is trying to compress its pricing tiers rather than simply pushing frontier performance upward — a signal that cost-per-intelligence, not just raw capability, is becoming the competitive battleground. Until Anthropic publishes concrete pricing and benchmark numbers, buyers should treat every capability claim in this announcement as vendor-asserted rather than independently verified.