Claude Opus 4.6: New AI flagship with 1 million tokens of context

What's new

On February 5, 2026, Anthropic released a new version of its top model, Claude Opus 4.6. The big news is a huge memory window of up to 1 million tokens (tokens are the AI's way of splitting text into small pieces, roughly three quarters of a word). It is the first time an Opus model can remember this much at once, so it can now work with entire codebases or long documents without losing track. The model can also decide for itself how much to think (adaptive thinking), and the developer tool Claude Code now has agent teams, where multiple AI agents work together on the same project.

What's clever

Anthropic calls it a shift to "vibe working": instead of steering the AI with precise commands step by step, you can now hand entire tasks over to it, such as a financial analysis or a legal review, and let it plan the way there on its own. In practice, this means fewer corrections afterward, because the model better understands what you are actually after and makes fewer messes along the way. There is also a Claude for PowerPoint in a test version, plus an upgraded Claude for Excel, so regular office workers can get help directly in the programs they know.

Cheaper or better?

The price in Anthropic's developer tool (API) is unchanged from its predecessor: $5 per million input tokens and $25 per million output tokens. If you use the very long conversations of up to 1 million tokens, it costs extra: $10 and $37.50 per million tokens. So Opus 4.6 is not cheaper, but you get more model for the same money. In tests, it beats both its own predecessors and the rival GPT-5.2 from OpenAI. On the GDPval-AA test, which measures how well an AI handles real work tasks, Opus 4.6 beats GPT-5.2 by around 144 points on an Elo scale (the same kind of rating system used in chess), and its own predecessor Opus 4.5 by a full 190 points. It also sets a record among Claude models on BigLaw Bench, a test of legal reasoning, with 90.2 percent.

What it's good at

Opus 4.6 is especially strong at programming in large, complex codebases, where it plans, debugs and does code review more reliably than before. It is also good at long, coherent tasks such as financial analysis and legal research, where it has to remember large amounts of information at once. For a business, this means you can put the model on an entire work task and rely more on the result without having to fix it afterward.