What Claude Opus 4.6 is
Claude Opus 4.6 is a large language model made by Anthropic — the AI safety company behind the Claude family of assistants. Think of it as a very capable AI that can read, write, reason, and write code. What set Opus 4.6 apart when it launched was its ability to handle enormous amounts of text at once and to carry out long, complex tasks on its own, checking in with you only when necessary.
Why it matters — the "why should I care" version
Most AI tools work best on short, focused tasks: answer this question, summarize this document, fix this bug. Opus 4.6 was designed to go much further. It can hold an entire large codebase in its "working memory" at once (thanks to a 1 million token context window — roughly 750,000 words), and it can plan and execute multi-step projects over hours without constant supervision. For developers, that means less babysitting and more getting things done.
The headline numbers
When it launched, Opus 4.6 claimed the top spot on four major AI benchmarks: Terminal-Bench 2.0 (a test of software engineering in real terminal environments), Humanity's Last Exam (a very hard knowledge test), GDPval-AA (a broad capability evaluation where it beat OpenAI's GPT-5.2 by 144 Elo points — a meaningful margin), and BrowseComp (a web-research task). These rankings are a snapshot in time — the AI field moves fast — but they confirmed Opus 4.6 as a frontier model at launch.
What "1 million tokens" actually means
A token is roughly three-quarters of a word. One million tokens is enough to hold about 750,000 words — or the equivalent of several full-length novels, a large software project, or months of email threads — all in one conversation. Before models like Opus 4.6, you'd have to chop big projects into pieces and feed them in separately, losing context along the way. With a million-token window, the AI can see the whole picture at once.
Agentic features: the AI that keeps working
Opus 4.6 shipped with two features that make long autonomous tasks practical:
- Agent teams in Claude Code: Claude Code (Anthropic's AI coding tool) can now spin up multiple AI "sub-agents" that work in parallel on different parts of a problem, then combine their results.
- Context compaction: When a task runs so long that even a million tokens isn't enough, the model intelligently summarizes earlier parts of the conversation so it can keep going without losing the thread.
Real-world proof: finding Firefox bugs
Shortly after launch, Anthropic partnered with Mozilla to test Opus 4.6 on a real security challenge: scanning Firefox's source code for vulnerabilities. Over two weeks, the model scanned nearly 6,000 code files and filed 112 reports. Mozilla classified 14 of the findings as high-severity — nearly a fifth of all high-severity Firefox vulnerabilities fixed in all of 2025. This wasn't a benchmark — it was a real software project with real stakes.
What it costs
Opus 4.6 kept the same price as its predecessor, Claude Opus 4.5: $5 per million tokens for input (what you send in) and $25 per million tokens for output (what the model sends back). For context, a million tokens is a lot — most everyday conversations use a few thousand at most.
Where it fits in the bigger picture
Opus 4.6 was part of a rapid release cadence at Anthropic. It followed Claude Opus 4.5 (March 2026) and was itself followed by Claude Opus 4.7 (May 2026), which added cybersecurity safeguards and vision improvements. The current Anthropic flagship is Claude Opus 4.8. Opus 4.6 also served as the foundation for Anthropic's Project Glasswing cybersecurity initiative, which used Mythos-class models (the next tier up) to scan critical infrastructure for vulnerabilities.
The bottom line
Claude Opus 4.6 was a meaningful moment in AI: it made million-token context and multi-hour autonomous tasks practical and affordable for the first time in the Claude line. It has since been superseded, but it set the template for what Anthropic's flagship models now do as a baseline.
