Claude Opus 4.8 is a large language model developed by Anthropic, released on May 28, 2026. Described as a "modest but tangible improvement" over its predecessor, Opus 4.7, the model focuses on enhancing reliability in autonomous engineering, long-horizon agentic coding, and complex reasoning tasks. It maintains the same pricing structure as previous Opus 4 series models while introducing refined internal logic for tool use and instruction following.
A central feature of this version is Adaptive Reasoning (also referred to as Adaptive Thinking), which enables the model to dynamically evaluate whether a specific turn in a conversation requires deep reasoning before responding. This bimodal approach allows for direct answers on simple lookups while applying extensive deliberation to multi-step problems, reducing token waste on simpler tasks. Additionally, the model introduces Effort Control, allowing users to manually toggle between effort levels—including High, Extra, and Max Effort—to prioritize either response speed or the depth of analytical rigor.
Agentic Capabilities and Honesty
The model demonstrates significant advancements in agentic performance, particularly in benchmarks such as the "Super-Agent" and "Legal Agent" tests. Technical evaluations indicate that Claude Opus 4.8 is approximately four times less likely than prior versions to allow bugs in its generated code to pass without flagging them. This focus on "honesty" allows the model to surface uncertainty more readily and push back against unsound user plans, making it more effective for high-stakes enterprise workflows and production-ready software engineering.
Technical Specifications
Claude Opus 4.8 supports a 1,000,000-token context window natively on the Claude API and major cloud platforms, with a maximum output limit of 128,000 tokens. It features a "Fast Mode" capability that operates at 2.5× the standard speed for latency-sensitive applications. The model also supports mid-conversation system messages, which allows developers to update instructions within a long-running session without re-sending the full prompt history, thereby optimizing prompt caching efficiency.