Claude Opus 4.6: What Builders Need to Know

The short version

Opus 4.6 is not a cosmetic bump. It's a model upgrade plus an execution layer upgrade.

The model itself improves coding, planning, and long-horizon agentic tasks. The platform adds Agent Teams, a compaction API, adaptive thinking, and effort controls. And the price stays the same: $5/$25 per million input/output tokens.

If you build with Claude today, this release changes what you can attempt in a single run-and how you structure teams around it.

What's new in Opus 4.6

1M token context window

Anthropic's Opus 4.6 is the first Opus-class model with a 1M token context window.

That's the headline. But the bigger story is what this enables: full-repo reasoning, multi-month project memory, and the ability to combine code, tickets, docs, and logs without aggressive pre-chunking.

Better at long-horizon agentic work

Anthropic is explicit here: Opus 4.6 improves long-horizon tasks and planning.

That matters if you're running multi-step operations like migrations, audits, or complex refactors where the model must hold state across hundreds of steps.

Stronger coding and planning

The release positions Opus 4.6 as a genuine upgrade to Opus 4.5 for coding and planning.

The data backs it: Opus 4.6 is SOTA on Terminal-Bench 2.0, Humanity's Last Exam, and BrowseComp. It beats GPT-5.2 by 144 Elo on GDPval-AA, and Opus 4.5 by 190 Elo. For a detailed comparison with OpenAI's latest, see our AI model wars breakdown.

Real-world productivity examples

Anthropic includes pragmatic metrics:

• One team closed 13 issues and assigned 12 in a single day across a 50-person org.
• A multi-million-line codebase migration completed in half the time.
• 38/40 cybersecurity investigations ranked best vs Opus 4.5.

These are not lab numbers. They're evidence that the "agentic" claims move the needle.

What's new in the product layer

Agent Teams in Claude Code

Agent Teams is the most important tooling addition in this release.

It allows multiple Claude Code instances to work together. A team lead coordinates, teammates execute. Teammates can message each other directly, which is a major difference from subagent-only systems.

You can run it in a split-pane mode (tmux/iTerm2) or in-process. There's a delegate mode to force the lead into coordination only, and a plan approval workflow for risky tasks.

Compaction API

Compaction lets you "compress" context while preserving key details.

In practice, this is a new option for long-running workflows: you can keep state without re-feeding full history. That shifts the cost curve and makes long-horizon agents more viable.

Adaptive thinking + effort controls

Adaptive thinking is a practical knob, not a marketing term.

You can steer how much reasoning the model applies to a task, and effort controls help manage both latency and spend. For builders who ship, this is the difference between "nice demo" and "production-ready."

Claude in Excel + Claude in PowerPoint

Excel improvements are incremental but important for analysis workflows. PowerPoint integration is new, research preview only.

If you build internal tooling, this hints at a deeper workspace integration strategy. It's still early, but it's a signal.

Benchmarks: what to trust

There are three benchmark lines that matter most here:

• Terminal-Bench 2.0: SOTA.
• GDPval-AA: +144 Elo vs GPT-5.2, +190 Elo vs Opus 4.5.
• BigLaw Bench: 90.2% (highest Claude score).

On finance, Anthropic reports a 23-point improvement over Sonnet 4.5. That's a notable jump for teams building analysis tools.

The main takeaway: this is not a marginal upgrade, and it hits both coding and professional knowledge work.

What actually changes for builders

You can collapse more steps into one run

The combination of 1M context + better planning means fewer orchestration layers.

Instead of splitting a task into eight "micro-runs," you can keep a single agent alive longer. That reduces glue code and failure points. This also reshapes how you approach multi-agent orchestration patterns-sometimes a single agent with enough context replaces a multi-agent pipeline entirely.

Multi-repo or mono-repo reasoning becomes practical

A 1M context window is big enough to hold large chunks of a mono-repo plus docs and tickets.

This enables a class of workflows that were previously painful: cross-layer changes, large refactors, or dependency audits without constant re-prompting.

The cost equation changes

Pricing is unchanged: $5/$25 per million input/output tokens.

With compaction and effort controls, you can hold state without constantly paying for full context. It's a more controllable spend profile for long jobs.

Agent Teams change how you structure work

Instead of one mega-prompt, you can split work into teammates.

That maps well to real-world engineering. You can assign a teammate to docs, another to tests, another to migration scripts, while the lead keeps coordination and risk control.

Practical use cases that now make sense

Large migrations

The "multi-million-line codebase migration in half the time" is the story to focus on.

If you've avoided refactors because the coordination overhead was too high, Opus 4.6 plus Agent Teams makes them feasible.

Security investigations

Anthropic highlights 38/40 cybersecurity investigations ranked best vs Opus 4.5.

This matters for incident response, log analysis, and internal security workflows where long context and precise reasoning are required.

Legal and finance analysis

BigLaw Bench at 90.2% and a 23-point finance improvement over Sonnet 4.5 are signals for professional workflows.

If you build for regulated industries, Opus 4.6 is shaping up as the safer pick for document-heavy tasks.

What this means for builders

• Plan bigger tasks. The 1M context and stronger long-horizon planning mean you can consolidate complex work into fewer runs.
• Adopt Agent Teams early. It matches how real teams work and gives you control knobs like delegate mode and plan approvals.
• Budget differently. Same pricing, but compaction and effort controls let you optimize spend without sacrificing depth.
• Use Opus 4.6 where trust matters. Legal, finance, and security work are trending stronger here than prior releases.

Trade-offs and open questions

Latency vs depth

Effort controls help, but bigger context usually means more compute.

You'll still need to decide which tasks need full reasoning depth and which should be cheap, fast passes.

Tooling maturity

Agent Teams is powerful, but new. You'll need guardrails, especially around plan approval and delegation workflows.

The benefits are obvious, but it's not a drop-in replacement for existing agent orchestration yet.

Office integrations are early

PowerPoint is still a research preview. It's promising, but not a reason to switch platforms.

Treat it as an early signal rather than a current capability.

Recommended adoption path

1. Pilot on migrations or audits. These are long-horizon tasks with clear success metrics.
2. Use Agent Teams for cross-layer changes. It mirrors how teams already work.
3. Test compaction + effort controls. Build a cost-performance profile for your own workloads.
4. Decide on office workflows later. PowerPoint is interesting, but not core yet.

Bottom line

Opus 4.6 is the most builder-friendly Claude release so far.

The raw model gains are real, but the bigger shift is in capability architecture: 1M context, Agent Teams, compaction, and effort controls make long-horizon work more realistic.

If you've been waiting for an LLM that can manage serious, multi‑day engineering tasks without constant babysitting, this is the first Opus release that feels ready for it. Our complete guide to building AI agents in 2026 walks through how to put these capabilities to work.