Anthropic launches Claude Opus 4.6 with 1M context

Anthropic has launched Claude Opus 4.6, an AI model designed for complex tasks that improves programming, analysis, and autonomous work. Its main new feature is a 1 million token context window in beta, enough to work with large amounts of code or documents without losing as many details along the way.

The model is already available on claude.ai, Anthropic's API, and major cloud platforms. Base pricing remains at $5 per million input tokens and $25 per million output tokens.

More capacity for long tasks

Opus 4.6 is designed to break complex work into steps, use tools, coordinate agents, and review its own results. In practice, it can analyze an extensive codebase, find the source of an error, propose changes, and check whether those changes work.

It also improves at research, financial analysis, and creating documents, spreadsheets, and presentations. Anthropic says the model pays more attention to difficult parts and moves faster through simple tasks, although that additional reasoning can increase time and cost for straightforward work.

The company recommends adjusting the /effort parameter when the model is spending too much time on a task. Four levels are available: low, medium, high, which is the default, and maximum.

The 1 million token context window, explained

The context window is the amount of information a model can keep in mind while it works. With 1 million tokens, Opus 4.6 can process very long documents, complete repositories, or extended conversations in a single session.

Anthropic says the model retains relevant information better as the context grows. In a test measuring the retrieval of hidden data within large volumes of text, it scored 76%, compared with 18.5% for Claude Sonnet 4.5. The feature is in beta, and premium pricing applies to requests exceeding 200,000 tokens: $10 per million input tokens and $37.50 per million output tokens.

The API also includes context compaction, a feature that automatically summarizes older parts of a conversation as it approaches the limit. This allows an agent to keep working for longer without having to start from scratch.

The figures Anthropic reports

According to the evaluations shared by the company, Claude Opus 4.6 leads in several programming, search, and reasoning tests. Its results include:

• The highest score on Terminal-Bench 2.0, a programming test using tools.
• First place among frontier models on Humanity’s Last Exam, a multidisciplinary reasoning evaluation.
• An advantage of around 144 Elo points over GPT-5.2 on GDPval-AA, a test of professional tasks in areas such as finance and law.
• An advantage of 190 Elo points over Claude Opus 4.5 on the same evaluation.
• The best result on BrowseComp, which measures the ability to find information that is difficult to locate online.

These figures come from tests described by Anthropic and do not mean the model will outperform other systems on every task. In addition, some evaluations use external tools, different levels of reasoning, or special configurations.

New features for programmers and teams

Claude Code, Anthropic's programming tool, now includes agent teams in research preview. Users can assign parts of a task to several agents operating in parallel, which is useful for reviewing different areas of a repository or investigating independent issues.

In the API, adaptive reasoning allows Claude to decide when it needs to think longer before responding. Developers can control this behavior to balance intelligence, speed, and cost.

Anthropic has also updated Claude in Excel so it can plan tasks, organize unstructured data, and carry out multistep changes. Claude in PowerPoint is available in research preview for Max, Team, and Enterprise plans, with features for creating presentations that preserve existing templates, fonts, and layouts.

More capacity also requires more control

Anthropic says Opus 4.6 maintains a safety profile equal to or better than Opus 4.5. The company tested behaviors such as deception, excessive flattery, cooperation with harmful uses, and responses that could cause harm.

The model also has stronger cybersecurity capabilities. This can help find and fix vulnerabilities, but it increases the risk of misuse. For that reason, Anthropic has created six new tests to detect dangerous responses and warns that it may add real-time interventions to block abuse.

For you, the most important change is not just that Claude responds better. It can work longer with less supervision, across more information, and on tasks that previously had to be divided manually. What remains to be watched is whether that autonomy holds up outside controlled tests and how much it ultimately costs when tasks actually use the full context window.