Anthropic Launches Claude Opus 4.6

Ronni Holmvig Strøm · 2026-02-05

Anthropic today unveiled Claude Opus 4.6, the latest upgrade to its flagship frontier model family.

Billed as the company's most capable system to date, Opus 4.6 delivers meaningful gains in long-horizon planning, autonomous task execution, large-scale code understanding, and professional knowledge work—while introducing a beta 1-million-token context window to the Opus line for the first time.

The release arrives less than three months after Claude Opus 4.5 (November 2025) and continues Anthropic's rapid iteration cadence in the race to build reliable, high-stakes AI agents.

Key Improvements and New Capabilities

Opus 4.6 focuses on reliability over raw scale. Anthropic highlights several qualitative advances:

Stronger agentic behavior — longer sustained task execution with fewer drop-offs or drift More deliberate planning and self-correction — improved code review, debugging, and mistake-catching loops Better performance on large codebases** — more consistent reasoning across tens or hundreds of files

A standout addition is the beta 1M token context window (up from 200K in the standard Opus 4.5/4.6 tier), enabling analysis of extremely long documents, full repositories, or multi-year financial/legal archives in a single pass. Maximum output tokens have also increased to 128K.

The model introduces "agent teams" (currently in research preview within Claude Code), allowing multiple Claude instances to divide complex projects—e.g., one agent owns frontend, another API logic, a third migration—and coordinate autonomously.

Benchmark Leadership

Anthropic reports state-of-the-art results across several difficult evaluations:

Highest score on Terminal-Bench 2.0 (agentic coding in realistic terminal environments) Leading position on Humanity’s Last Exam (multidisciplinary hard-reasoning test) Largest margin on GDPval-AA — an economic-value knowledge-work benchmark spanning finance, legal, and strategy — where Opus 4.6 outperforms OpenAI's GPT-5.2 by ~144 Elo points and its own predecessor (Opus 4.5) by ~190 points Top result on BrowseComp (ability to find obscure but critical information online)

Early demonstrations include Opus 4.6 (using agent teams) autonomously building a from-scratch, optimizing C compiler capable of handling substantial codebases—something previous Opus generations could not reliably achieve.