Claude Opus 4.6: 1M context, effort levels and agent teams in Claude Code

Anthropic released Claude Opus 4.6 on February 5. It is not a generational leap, but the list of changes is large for a “point” release: a 1 million token context window in beta, up to 128k output tokens, explicit control over how much the model reasons, and features aimed at long-running agents. Since I use Claude every day for programming, I read the announcement with an eye on what changes in practice.

Access, context and pricing

The model is on the API as claude-opus-4-6, as well as on claude.ai and the major cloud platforms. The key points:

1M-token context in beta, available only on the Claude Developer Platform. According to Anthropic, it is the first Opus-class model with 1M context.
Up to 128k output tokens.
Standard pricing of $5 per million input tokens and $25 per million output tokens.
Prompts over 200k tokens cost $10 input and $37.50 output per million, on the Claude Platform only.
US-only inference costs 1.1x the per-token price.

The over-200k pricing detail matters: using the full window is not just “more tokens”, it is a different price tier. For tasks that fit in 200k, nothing changes; for those that do not, it is worth comparing against context-retrieval strategies before dumping everything into the prompt.

Controlling reasoning: adaptive thinking and effort levels

Opus 4.6 brings adaptive thinking: the model decides on its own when deeper reasoning helps. For those who want control, there are effort levels low, medium, high (the default) and max.

This solves a real pain. In pipelines with many simple calls, paying for a large model’s reasoning on every one is wasteful; on hard tasks, you want the opposite. An explicit knob lets you use the same model in both scenarios, tuning cost and latency per call.

Context compaction also arrives in beta, summarizing and replacing older context as a conversation nears a configurable threshold. For agents that run for hours, it is the difference between stopping at the window limit and continuing to work.

Agents: agent teams and Claude in PowerPoint

In Claude Code, the new feature is agent teams, in research preview: multiple agents work in parallel and coordinate on their own. It is a natural direction for large tasks, such as refactors that touch many files or investigations that can be split up.

Outside the dev world, Claude in PowerPoint enters research preview for Max, Team and Enterprise plans.

The numbers, according to Anthropic

The results Anthropic itself published:

Terminal-Bench 2.0: the highest score among the models evaluated on this agentic coding benchmark.
GDPval-AA: about 144 Elo points ahead of GPT-5.2 and 190 ahead of Opus 4.5.
MRCR v2 (8-needle, 1M): 76%, versus 18.5% for Sonnet 4.5.
BrowseComp: 86.8% with a multi-agent harness.

The MRCR result is the most relevant one for anyone thinking about using the 1M context. A large window is only valuable if the model can find and use information spread across it, and the jump over Sonnet 4.5 suggests the window is usable, not just a marketing number.

On safety, Anthropic reports low rates of misaligned behavior in its automated audit, similar to Opus 4.5, and the lowest over-refusal rate among recent Claude models. It also added new cybersecurity probes to detect misuse.

Caveats

The 1M context is beta and exists only on the Developer Platform. If you use the model through other platforms, check availability before planning around it.
Agent teams and Claude in PowerPoint are research previews: expect changes.
All benchmarks come from Anthropic; evaluate on your own code and your own tasks.

My take

What interests me most in Opus 4.6 is not any single benchmark but the combination of controls: effort levels, adaptive thinking and context compaction. Together, they make it viable to use a flagship model in long agent workflows without costs spiraling. I would start by treating effort level as a product parameter (cheap by default, max only when needed) and treat the 1M context as a tool for specific cases, not the new normal, until I understand how the over-200k price tier affects the bill.