Claude Opus 4.6 Is Here: Agent Teams, 1M Context, and the Most Agentic Model Yet

Claude Opus 4.6: The Biggest Claude Code Update Yet

Anthropic released Claude Opus 4.6 today, and it’s more than just a model upgrade — it comes with two game-changing features for Claude Code users: Agent Teams and Effort Tuning.

What’s New

1. Claude Opus 4.6 Model

The new model is fundamentally more capable:

Better planning: Deliberates more carefully before acting, reducing wasted steps
Longer agentic sessions: Sustains focus across extended multi-step operations
Large codebase reliability: Works effectively in massive, real-world codebases
Self-correction: Catches its own mistakes during code review and debugging
1M token context (beta): First Opus-class model with million-token context window — 76% accuracy on MRCR v2’s 8-needle 1M test vs Sonnet 4.5’s 18.5%

Pricing stays the same: $5/$25 per million input/output tokens.

2. Agent Teams (Research Preview)

The headline feature for Claude Code. Instead of one agent working sequentially, you can now orchestrate a team of Claude Code instances working in parallel:

A lead agent coordinates work and spawns teammates
Teammates work independently with their own context windows
Teammates can message each other directly — not just report back
A shared task list tracks dependencies and auto-unblocks

Use cases: Parallel code review, competing-hypothesis debugging, cross-layer feature development, research tasks.

Enable it:

// settings.json
{ "env": { "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1" } }

3. Effort Tuning

Control how much the model thinks. Run /model and use arrow left/right:

Less effort = Faster, cheaper
More effort = Better results for complex tasks

Benchmark Highlights

Opus 4.6 leads or matches the best across nearly every benchmark:

Area	Result	vs Competition
ARC AGI 2	68.8%	Nearly 2x Opus 4.5 (37.6%)
Terminal-Bench 2.0	65.4%	Highest score (beats GPT-5.2 Codex CLI)
BrowseComp	84.0%	+24% ahead of nearest competitor
GDPVal-AA	1606 Elo	+144 points vs GPT-5.2
Humanity’s Last Exam	53.1%	Highest with tools
t2-bench Telecom	99.3%	Highest agentic tool use

Also 2x better than Opus 4.5 on computational biology, organic chemistry, and phylogenetics. 90.2% on BigLaw Bench with 40% perfect scores.

What This Means for You

Longer, more reliable coding sessions — the model won’t degrade mid-task
True parallelism with Agent Teams — specialists working simultaneously and communicating
Token savings with Effort Tuning — dial down for simple tasks, dial up for complex ones
Fewer “I forgot” moments — 1M context means massive codebases stay in memory
Better self-correction — catches bugs before you need to point them out

Also Released

Claude in PowerPoint (research preview): Generates presentations from descriptions or templates
Claude in Excel: Improved long-running task handling and multi-step changes
Context Compaction API (beta): Auto-summarizes older context for long agentic operations
128k output tokens now supported

Quick Start

# Update Claude Code to latest
claude update

# Enable Agent Teams
# Add to settings.json: { "env": { "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1" } }

# Try effort tuning
# Run /model and use arrow keys to adjust

This newsletter was written by Claude Opus 4.6 itself. Meta enough for you?