Anthropic: Claude Opus 5 (batch)

Claude Opus 5 (batch) is Anthropic's Opus-tier flagship exposed through the Message Batches API, where workloads are processed asynchronously at roughly half the standard input and output token cost. It is the same underlying model as claude-opus-5: a large transformer with adaptive thinking enabled by default, a one-million-token context window, and strong multimodal understanding over text and images. The batch endpoint suits high-volume offline jobs such as evaluation sweeps, dataset labeling, and large document or codebase analysis that do not require real-time latency.

Anthropic positions Opus 5 as a major step over Opus 4.8 in agentic coding, computer use, and long-horizon knowledge work, often approaching Claude Fable 5 intelligence at lower Opus pricing. Reported strengths include software engineering (SWE-bench family), terminal and repository agents, web research (BrowseComp), professional task quality (GDPval-AA), and graduate-level reasoning (GPQA Diamond, Humanity's Last Exam). The July 2026 system card and launch materials document extensive safety, alignment, and capability evaluations across cyber, biology, and agentic misuse settings.

Developers invoke the model via claude-opus-5 on the Batches API across Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry. Standard synchronous calls support up to 128K output tokens; batch jobs can reach up to 300K output tokens with the output-300k-2026-03-24 beta header. Knowledge cutoff is May 2026. Claude Opus 5 remains available but Anthropic recommends newer Opus 5.5 for improved performance on most tasks.