s the industry anxiously awaits the release of Claude Sonnet 5, Anthropic released Claude Opus 4.6 to match OpenAI's short-lived advantage in agents.
On Thursday, Anthropic released the latest update to its Claude model family, saying the model offers better coding and review skills, improved task planning, and can sustain agentic tasks for longer than its predecessor.
Alongside the launch, Anthropic unveiled a feature called “agent teams,” which gives users the ability to spin up agents that can split up tasks autonomously and work on them in parallel. Notably, this feature offers capabilities similar to OpenAI’s Codex, which debuted multi-agent capabilities earlier this week.
Claude Code has been a viral hit over the past couple months. It's even been blamed for the stock market sell-off of software companies because investors worry that they may not have a future if any organization or individual can now use AI to vibe-code custom software that exactly meets their needs.
However, OpenAI's Codex raised the bar. Rather than just offering an AI agent for coding, it enabled the ability to use a team of agents working together. With Opus 4.6, Claude can now do the same.
Anthropic also touted a number of other improvements that Opus 4.6 has to offer, including improved abilities on work tasks such as financial analysis, research, document creation and agentic search; the ability to work more reliably with large codebases; and autonomous multitasking capabilities when utilized in Claude Cowork.
- The model is also the first in the Opus line to offer a 1 million token context window,
- Opus 4.6 also achieves state-of-the-art performance on several benchmarks, including agentic coding evaluation Terminal-Bench 2.0, frontier model evaluation Humanity’s Last Exam, and finance and legal evaluation GDPval-AA.
- One drawback that the company noted in its blog post is that, while Opus 4.6 is more thoughtful and careful in considering its outputs to offer better results for harder problems, this feature can “add cost and latency on simpler ones.”
Alex Albert, head of Claude relations at Anthropic, said in an article on X that the launch represents “the watershed moment for AI becoming a real working partner for people who spend their days in spreadsheets, slide decks, and long docs.”
The Deep View got access to the model before the launch, and did not see an observable jump in the quality of the outputs. Having said this, the previous model was already very capable, and this release did not hinder that experience at all. It is possible that you would only see the difference when stress testing the model under really complex coding and reasoning workflows, and our team will be continuing to test it to find the latest nuances.