What is Claude Opus 4.6, and what are agent teams and adaptive thinking?

Yesterday, Thursday, February 5, 2026, Anthropic released Claude Opus 4.6, the newest version of its top model. TechCrunch notes that Opus 4.5 came out only last November, so this is a quick follow-up rather than a new generation.

Anthropic's message is that Opus, long valued mainly by programmers, is now meant for a wider range of knowledge workers too. Scott White, Anthropic's head of product, told TechCrunch that many people who are not professional developers had been using Claude Code simply because it gets tasks done, including product managers and financial analysts.

What Anthropic released

Opus 4.6 is available on claude.ai, through the API as claude-opus-4-6, and on the major cloud platforms. The price is unchanged: $5 per million input tokens and $25 per million output tokens. A token is a small piece of text, often part of a word.

Anthropic says the model:

• plans more carefully and spends its effort on the hard parts of a task;
• keeps going longer on agentic tasks, meaning multi-step jobs where it uses tools on its own;
• handles larger codebases and catches more of its own mistakes in code review and debugging;
• does better at financial analysis, research and working with documents, spreadsheets and slides.

It also reports leading scores on Terminal-Bench 2.0 (coding in a command-line terminal), Humanity's Last Exam (hard expert questions) and BrowseComp (finding obscure facts online). On GDPval-AA, a test of professional tasks in fields such as finance and law, it says Opus 4.6 is about 144 Elo points ahead of OpenAI's GPT-5.2, a gap that translates into winning roughly 70% of head-to-head comparisons. These are Anthropic's own figures.

A much longer memory: the 1M context window

The context window is how much text a model can consider at once, its working memory for a conversation. Opus 4.6 is the first Opus model with a 1-million-token window, enough for a large codebase or a stack of long reports. It is in beta on Anthropic's developer platform, and prompts over 200,000 tokens cost more: $10 per million input tokens and $37.50 per million output tokens.

A big window is only useful if the model can still find things in it. Models often suffer from "context rot," getting worse as a conversation grows. Anthropic tested this with MRCR v2, which hides eight pieces of information in a million tokens and asks the model to retrieve them:

Opus 4.6 can also write answers of up to 128,000 tokens, so long documents or large code changes need fewer separate requests.

How the new features work

Adaptive thinking. Until now, developers simply turned extended thinking (step-by-step reasoning before answering) on or off. Now the model judges from the request how much thinking is worth doing.

Effort levels. Developers choose low, medium, high (the default) or max. Higher effort means deeper reasoning but more time and cost. Anthropic admits Opus 4.6 sometimes overthinks easy tasks and suggests dropping to medium when that happens.

Context compaction (beta). When a long task nears the window's limit, Claude automatically summarizes older parts of the conversation and replaces them, so it can keep working.

Agent teams. In Claude Code, a research preview lets you launch several agents that work in parallel and coordinate among themselves, instead of one agent doing everything in sequence. Anthropic suggests them for work that splits into independent, mostly reading tasks, such as reviewing a codebase. You can step into any agent's work directly.

Office tools. Claude in Excel gets stronger at long, multi-step tasks and can work out structure from messy data. Claude in PowerPoint arrives as a research preview for subscribers on the Max, Team and Enterprise plans. TechCrunch explains the difference: before, Claude could produce a deck that you then opened and edited in PowerPoint; now it works in a side panel inside PowerPoint, reading your layouts, fonts and slide masters.

Safety

Anthropic says Opus 4.6 is as well aligned as Opus 4.5, with low rates of deception or sycophancy (telling people what they want to hear), and the lowest rate of over-refusals, turning down harmless requests, among its recent models. Because the model is strong at finding software vulnerabilities, Anthropic added six new cybersecurity probes, tools that detect harmful uses, and says it may introduce real-time blocking of abuse.

How do you choose the right effort level for Claude Opus 4.6?

Effort is a dial between speed and depth. High, the default, lets the model think whenever it judges thinking useful, and suits most real work: debugging, analysis, writing with several constraints. Max is for the hardest problems, where a better answer is worth waiting and paying for. Medium or low fit quick, simple requests such as reformatting text, short summaries or routine code edits, where deep reasoning only adds delay.

A practical way to decide is to run the same task at two levels and compare:

Review this function for bugs and edge cases. List each problem,
why it matters and a suggested fix. If you find nothing serious,
say so in one sentence.

If medium finds the same problems as high, keep medium for that kind of task. If high catches something medium missed, the extra time is earning its keep. Effort is set through the API, so this dial matters most to developers building with Opus 4.6.