# Claude Opus 4.6 in 10 Minutes Source: https://www.youtube.com/watch?v=r2zxcB67vwM Recap page: https://rapidrecap.app/video/r2zxcB67vwM Generated: 2026-02-09T08:03:48.246+00:00 --- ## Quick Overview Anthropic announced Claude Opus 4.6, which significantly improves coding skills, agentic capabilities, and long-context handling, achieving state-of-the-art results on several benchmarks, including leading frontier models on agentic coding evaluations like Terminal-Bench 2.0, while also introducing features like agent teams and context compaction. **Key Points:** - Claude Opus 4.6 significantly improves upon its predecessor's coding skills, enabling more reliable operation in larger codebases and better code review/debugging. - Opus 4.6 features a 1M token context window in beta and excels at finding information across long contexts, notably scoring 93.0% on the MBRR long-context retrieval benchmark, a substantial leap over Opus 4.5's 76.0%. - The model achieves the highest score on the agentic coding evaluation Terminal-Bench 2.0, outperforming all other frontier models listed, including GPT-4.5, Sonnet 4.5, and Gemini 3 Pro. - New API features include the ability to assemble agent teams for complex tasks, utilizing main orchestrator agents and subagents that can coordinate directly or through shared resources. - The model introduces context compaction, allowing it to summarize large contexts without losing critical details, and adaptive hinting, where the model self-regulates its reasoning depth based on task complexity. - The new model also shows strong performance in agentic reasoning benchmarks, scoring 73.8% on GraphWalks long-context reasoning, significantly better than Opus 4.5's 36.7%. - Anthropic also released upgrades to Claude in Excel and Claude in PowerPoint as part of this announcement. ![Screenshot at 00:00: The Anthropic announcement page for Claude Opus 4.6 is displayed, introducing the model and its key improvements in coding and agentic capabilities.](https://ss.rapidrecap.app/screens/r2zxcB67vwM/00-00-00.jpg) **Context:** This video details the announcement of Claude Opus 4.6 by Anthropic, their latest and most capable model as of February 5, 2024. The presentation focuses heavily on the model's performance improvements across agentic tasks, coding benchmarks, and its vastly expanded context window capabilities, highlighting new features like agent teams and context compaction designed to enhance complex workflow execution. ## Detailed Analysis Anthropic introduced Claude Opus 4.6, positioning it as their smartest model yet, with major upgrades centered on coding, agentic abilities, and long-context understanding. In coding, Opus 4.6 shows substantial improvement over Opus 4.5, particularly on the Agentic Coding (DBRX Search) benchmark (80.0% vs 80.9% for Opus 4.5, but Opus 4.6 shows a much larger lead over competitors). It scored the highest on Agentic Terminal Coding (66.4%) and Agentic Gesture Use (99.3%). The long-context retrieval capability saw a massive leap, with Opus 4.6 achieving 93.0% on the MBRR benchmark compared to Opus 4.5's 76.0%, demonstrating superior ability to retrieve information from massive context windows (1M tokens in beta). Performance across general agentic evaluations is strong, with Opus 4.6 leading on Terminal-Bench 2.0 and showing large margins over other models like GPT-4.5 and Sonnet 4.5. New API features include agent teams, allowing users to orchestrate multiple Claude instances—one main agent coordinating multiple subagents—which can work on complex tasks like building a C compiler from scratch, a project that took 2,000 sessions and $20,000 in API costs. New concepts like context compaction and adaptive hinting are introduced to manage long context and tailor reasoning depth. The cost for Opus 4.6 via the API remains the same as Opus 4.5: $5/$25 per million tokens for input/output. ### Performance Benchmarks - Opus 4.6 achieved the highest score on Agentic Coding (80.0%) and Terminal Coding (66.4%) among tested models; it shows a substantial leap in long-context retrieval (93.0%) over Opus 4.5 (76.0%). ### Agent Teams & Orchestration - Users can now assemble agent teams where a main orchestrator agent coordinates multiple subagents, allowing for complex, multi-step workflows and parallel task execution, similar to organizing human teams. ### New API Features - Introduces context compaction to efficiently summarize large inputs and adaptive hinting to control the model's reasoning depth based on task complexity. ### C Compiler Project - A proof-of-concept where 16 agents built a functional C compiler from scratch, costing $20,000 in API calls over 2,000 sessions, demonstrating high-level autonomous software development capability. ### Limitations and Learnings - The compiler project revealed limitations in context window pollution and time blindness in LLMs, which necessitated designing specific harnesses to maintain progress and avoid excessive token usage. ### Model Availability and Pricing - Claude Opus 4.6 is available on claude.ai, the API, and major cloud platforms, with API pricing remaining $5/$25 per million tokens (input/output). ![Screenshot at 00:03: The blog post title announcing the release of Claude Opus 4.6 is prominently displayed.](https://ss.rapidrecap.app/screens/r2zxcB67vwM/00-00-03.jpg) ![Screenshot at 00:40: A bar chart comparing Opus 4.6 against competitors on Agentic Coding \(Terminal-Bench 2.0\), showing Opus 4.6 achieving the highest accuracy.](https://ss.rapidrecap.app/screens/r2zxcB67vwM/00-00-40.jpg) ![Screenshot at 03:09: A bar chart illustrating the significant improvement in Long-Context Retrieval \(MBRGR v2.2\) for Opus 4.6 \(93.0%\) compared to Opus 4.5 \(76.0%\).](https://ss.rapidrecap.app/screens/r2zxcB67vwM/00-03-09.jpg) ![Screenshot at 05:58: Documentation detailing the new concept of Agent Teams, showing configuration settings for enabling experimental features.](https://ss.rapidrecap.app/screens/r2zxcB67vwM/00-05-58.jpg) ![Screenshot at 06:30: A terminal window screenshot indicating the agent team completed the task, showing the text "Then walked away" after building a C Compiler.](https://ss.rapidrecap.app/screens/r2zxcB67vwM/00-06-30.jpg)