Anthropic released Claude Opus 4.6 on February 5, 2026 — about ten weeks after Opus 4.5 set the bar for AI coding. The headline change is a 1M token context window in beta, but the more interesting story is the new “agent teams” coordination model and how Anthropic is now charging for both.

What Actually Changed

1M token context window (beta). Five times what Opus 4.5 offered. This puts Anthropic into the same long-context regime as Gemini 3 Pro for the first time.

Agent teams. Multiple Claude instances can now split a larger task into segments and coordinate in parallel. Anthropic frames this as one model orchestrating others — a foreman-and-crew pattern rather than a single agent grinding through a queue.

Adaptive thinking. The model now decides for itself how much extended thinking to apply, based on contextual cues from the prompt. Less of a knob you have to tune, more of something the model figures out.

Effort controls. Developers get explicit dials for intelligence / speed / cost trade-offs. If you know your workload is simple, you can tell the model to spend less compute on it.

Same pricing. $5 / $25 per million input/output tokens — unchanged from Opus 4.5. That is the move worth paying attention to.

Why the Pricing Decision Matters

Anthropic could have raised prices for the 1M context tier. They did not. The same dollar-per-token rate that bought you 200K context in November now buys you 1M context.

That is a meaningful shift in how to think about long-context workloads. Previously, feeding a model a whole codebase or a long meeting transcript was a “what’s the cheapest model that can swallow this” decision. With Opus 4.6 at flat pricing, you can keep using the frontier model and just give it more.

The catch is that 1M tokens at $5/M input is still $5 per query in the worst case. The economics flip from “can the model handle this” to “do I really need this much context, or am I being lazy.”

Agent Teams Is the Sleeper Feature

Single-agent loops have a clear limit: the model can only think about one thing at a time. For tasks that decompose cleanly — research a list of 50 candidates, summarize 100 documents, run 20 independent experiments — running them in series is wasteful.

Agent teams formalizes the pattern. One Claude instance plans and decomposes; others execute sub-tasks in parallel; a coordinator gathers results. You could already build this manually with the API. What is new is that Anthropic is treating it as a first-class workflow and presumably optimizing the model’s behavior around it.

For consultancy work specifically — where a typical engagement involves “look at twelve client systems and produce a comparative report” — agent teams is the feature I expect to use most.

Adaptive Thinking Reduces a Tuning Knob

In Opus 4.5, you set a thinking budget and lived with it. Too low and the model missed obvious things on hard prompts. Too high and you paid for thinking on prompts that did not need it.

Opus 4.6 reads the contextual signals — prompt complexity, presence of code, presence of constraints — and adjusts the thinking budget itself. This is the kind of feature that feels small in a press release and saves you fifteen minutes per week in practice.

The explicit effort-controls knob is still there if you want to override. The defaults are now better.

Where Opus 4.6 Fits

The pattern Anthropic is settling into:

  • Opus 4.5 (Nov 2025) — proved the coding lead with 80.9% SWE-bench
  • Opus 4.6 (Feb 2026) — extended the context and added orchestration primitives
  • Opus 4.7 (next) — presumably continues the trajectory

For teams already on Opus 4.5: the upgrade is free in the sense that pricing is unchanged. The 1M context is in beta, so production workloads might wait, but for exploratory work and agent loops, the new model is strictly more capable.

For teams choosing between Opus 4.6 and Gemini 3 Pro: both have 1M context now. The choice comes down to coding (Opus) versus multimodal reasoning (Gemini). The gap on each axis is smaller than the gap on the other.

Availability

Claude Opus 4.6 is available on:

  • claude.ai (Pro and Team plans)
  • The Anthropic API
  • AWS Bedrock, Google Cloud Vertex AI, Microsoft Azure Foundry

The 1M context window is in public beta on the API; you opt in per request.