On Monday, Anthropic unveiled Opus 4.5, the latest iteration of its flagship AI model. This release marks the final addition to the Anthropic 4.5 series, following the launches of Sonnet 4.5 in September and Haiku 4.5 in October.
As anticipated, the new Opus version excels across a range of benchmark evaluations, including coding assessments (SWE-Bench and Terminal-bench), tool utilization (tau2-bench and MCP Atlas), and general problem-solving tasks (ARC-AGI 2 and GPQA Diamond).
Notably, Opus 4.5 is the first model to surpass an 80% score on the SWE-Bench validation—a widely respected benchmark for evaluating coding capabilities.
Anthropic also highlighted Opus’s enhanced performance in computer interaction and spreadsheet functionalities, introducing complementary products to demonstrate these capabilities. Alongside Opus 4.5, the company is expanding access to Claude for Chrome and Claude for Excel—both previously in pilot phases. The Chrome extension will now be available to all Max-tier users, while the Excel-focused model will be rolled out to Max, Team, and Enterprise subscribers.
Opus 4.5 also features significant improvements in memory management for long-context operations, requiring substantial architectural adjustments to how the model handles memory.
“We’ve improved the overall quality of long-context handling in Opus 4.5’s training, but a larger context window alone isn’t sufficient,” Dianne Na Penn, Head of Research Product Management at Anthropic, told TechCrunch. “Knowing which details to retain is crucial—and that capability complements an extended context window.”
These enhancements also enable the long-requested “endless chat” feature for paid Claude users. When the model reaches its context limit, conversations can continue seamlessly without interruption. Instead of alerting the user, the model automatically compresses its contextual memory in the background.
Many of the upgrades are specifically tailored for agent-based workflows—particularly scenarios where Opus acts as a central orchestrator managing a team of Haiku-powered sub-agents. Executing such tasks demands robust working memory, precisely where the memory improvements described by Penn prove most impactful.
“This is exactly where foundational elements like memory become critical,” Penn explained. “Claude needs to navigate large codebases and extensive documents while also knowing when to backtrack and re-examine specific information.”
