Claude Opus 4.6: Coding Claims, Agent Evaluation, and a Longer Context

SUMMARY

In 30 seconds

Matthew Berman shares his early impressions of Claude Opus 4.6 and summarizes Anthropic’s claims about coding and longer agentic tasks. He also shows a task-horizon chart and mentions the model’s one-million-token context window, which Anthropic announced in beta at launch. The chart reflects performance on a specified evaluation, not a general guarantee of autonomous work.

1Early impressionBerman says his early access to Claude Opus 4.6 left him impressed relative to Opus 4.5.
2Anthropic’s coding claimsThe launch post describes better planning, work on larger codebases, review, and debugging.
3A chart of task horizonsThe displayed graph estimates the human task duration at which models have a 50% chance of success on its software-task evaluation.
4A larger context windowAnthropic announced a one-million-token context window in beta at launch.

Takeaways

Berman reports positive early impressions of Claude Opus 4.6 compared with Opus 4.5.
Anthropic claims improvements in planning, coding, review, debugging, and longer agentic tasks.
The chart’s task horizon describes a 50% success estimate on a particular set of software tasks, not a universal autonomous runtime.
Anthropic announced a one-million-token context window in beta at launch.

TopicsFoundation Models Coding agents Evals Context engineering

FULL TEXT

Read the edited text by chapter

Show transcript

Transcript only

Opus 4.6 and coding

Berman shares his positive early impression and recaps Anthropic’s launch claims. He highlights planning, longer tasks, and code review.

Claude Opus 4.6 is here. I had early access, and based on my experience I consider it a significant step forward over Opus 4.5. Anthropic’s [launch announcement](https://www.anthropic.com/news/claude-opus-4-6) says it plans more carefully, handles longer agentic tasks and larger codebases more reliably, and improves code review and debugging. Those are the company’s claims and my early impressions, not a complete independent evaluation.

Task horizons and context

He shows a chart of model performance on software tasks. He also points to the one-million-token context window announced in beta.

SLIDE 2 / 2Software-task horizon chartThe chart labels its measure as the task duration corresponding to an estimated 50% success rate.

I show a chart illustrating progress in model task horizons and describe the rise as nearly vertical. The graph labels the horizon as the human task duration for which its model predicts a 50% chance of successful completion on the evaluated software tasks. It is a measure tied to that evaluation and its setup, not a guarantee that an agent can run any task autonomously for that long.

Anthropic also announced a one-million-token context window in beta at launch. I describe that as a milestone and compare it with offerings from other companies at the time. Context-window size, however, is not itself a guarantee that all information within it will be used reliably; availability and product terms can change.