Home / AI

Claude Opus 4.6 Adds Agent Teams and a 1M-Token Window, Doubling Down on ARC-AGI-2

Anthropic's February flagship lets Claude Code spin up parallel sub-agents and scores 68.8% on a reasoning test its predecessor managed 37.6% on.

AIFeb 5, 20264 min read2 diagrams

Key takeaways

  • Opus 4.6 scored 68.8% on ARC-AGI-2 versus 37.6% for Opus 4.5 and 54.2% for GPT-5.2.
  • Claude Code gains 'agent teams', parallel sub-agents whose work is reconciled by a lead agent.
  • A 1M-token context window arrives in beta with context compaction; pricing stays at $5 and $25 per million tokens.

Anthropic released Claude Opus 4.6 on February 5, 2026, three months after Opus 4.5. The update is built around long-running agentic work: a million-token context window, automatic context compaction, and a Claude Code feature that lets several agents tackle a codebase in parallel.

What changed

The headline capability is "agent teams" in Claude Code. Instead of one agent working through a task serially, a lead agent can split work across sub-agents that run at the same time on different parts of a repository, then reconcile what comes back. It mirrors how human engineering teams divide a refactor, and it addresses a practical bottleneck: single-agent sessions are slow and run out of context.

The context story supports that. Opus 4.6 offers a 1M-token window, described as beta in some coverage, plus context compaction, which summarizes older material automatically as a session approaches its limit. Together they let an agent keep working on a large project for hours without a human clearing and restarting it. Pricing stayed at $5 per million input tokens and $25 per million output tokens.

Architecture diagram: a lead agent splits a task into three parallel sub-agents, then merges their results Lead agent splits the task Sub-agent A module 1 Sub-agent B module 2 Sub-agent C tests Reconcile and merge
The agent-teams pattern in Claude Code, simplified. Illustrative.

Benchmarks

On ARC-AGI-2, Opus 4.6 scored 68.8%, compared with 37.6% for Opus 4.5 and 54.2% for GPT-5.2, which coverage described as the largest single-generation jump on that test. On Terminal-Bench 2.0, which measures command-line agent tasks, it posted 65.4% against 59.8% for Opus 4.5, with one comparison table listing GPT-5.2 at 64.7%. As usual the figures are vendor-reported and based on secondary summaries of Anthropic's launch.

Bar chart of ARC-AGI-2: Claude Opus 4.6 68.8, GPT-5.2 54.2, Claude Opus 4.5 37.6 ARC-AGI-2 score (%) Claude Opus 4.6 68.8 GPT-5.2 54.2 Claude Opus 4.5 37.6 Vendor-reported at launch.
ARC-AGI-2 results as listed in launch coverage. Source: Philipp Dubach, OfficeChai.

Why it matters

The release shows where the frontier is competing. Raw question-answering benchmarks are saturating; the differentiators are now how long an agent can run, how much it can hold in memory, and how many copies of it can cooperate. Parallel agents multiply token use, which is good for revenue per task and a reason Anthropic kept per-token prices flat. Google answered two weeks later with Gemini 3.1 Pro, which reported 77.1% on ARC-AGI-2, so the lead lasted a short time.

The mechanics of long-running agents

Long context and compaction solve different problems. A bigger window lets a model see more of a repository, a contract set or a research folder at once, but attention over a million tokens is expensive and quality can degrade in the middle of the input. Compaction attacks the failure from the other side: as a session grows, the system condenses earlier turns into a shorter summary so the agent keeps the decisions and drops the chatter. The trade-off is that summaries lose detail, and an agent that forgets a constraint stated three hours ago may quietly break it.

Parallelism raises its own questions. Splitting a refactor across sub-agents is easy when the modules are independent and awkward when they share interfaces. The lead agent has to define contracts up front, collect the pieces, and run tests to catch conflicts. In human teams, that coordination overhead is what limits how many people can usefully work on one codebase, and the same law applies to agents: past a point, adding more of them produces more merge work than progress. Well-bounded changes are therefore the natural fit for the feature.

Cost behaves differently from single-agent runs. Each sub-agent carries its own context, so three agents reading the same large files consume roughly three times the input tokens. At $5 per million input tokens, a big parallel run can cost real money, which is why per-task cost, not list price, is the number buyers should track. It also explains why vendors push caching and compaction: they keep repeated context from being billed at full rate again and again.

The competitive backdrop is tight. GPT-5.2 had arrived in December with strong reasoning scores, and Gemini 3 Pro had set leaderboard marks in November. Opus 4.6 claimed first place on Terminal-Bench 2.0 at launch, then watched Google reclaim the ARC-AGI-2 lead within two weeks. For customers, the practical result is that the best model for a given job changes every few months, which favours tools and contracts that make switching cheap.

What to watch

  • Whether agent teams produce mergeable code in practice or mostly merge conflicts.
  • Cost per completed task, not per token, as parallel agents inflate usage.
  • The 1M window leaving beta and how compaction affects accuracy on very long sessions.

Sources

  1. Anthropic Releases Claude Opus 4.6 with 1M Token Context Window — NYU Shanghai RITS
  2. Claude Opus 4.6: Anthropic's New Flagship AI Model for Agentic Coding — Philipp Dubach
  3. Anthropic launches Claude Opus 4.6 with performance upgrades and coding accuracy — Business Today
  4. Anthropic Releases Claude Opus 4.6, Beats Gemini 3 And GPT 5.2 On Most Benchmarks — OfficeChai

More in AI