Claude Opus 4.6 for Long-Running Agents: Adaptive Thinking, Effort, Compaction and Agent Teams

Bottom line: The most practical changes in Claude Opus 4.6 are not a single benchmark score but the controls for long-running work: adaptive thinking, four effort levels, context compaction, up to 128K output tokens, and agent teams in Claude Code. Used well, they let one model handle both quick answers and multi-hour tasks without wasting tokens. All details below come from Anthropic’s announcement and documentation.

Why “autonomy” is the theme

Anthropic describes Opus 4.6 as a model that “brings more focus to the most challenging parts of a task without being told to, moves quickly through the more straightforward parts” and “stays productive over longer sessions.” Early-access partners quoted in the announcement describe it breaking complex tasks into subtasks and running tools and subagents in parallel. These are vendor and partner statements, not independent measurements, but they explain what the new controls are for.

Adaptive thinking

Previously, developers chose to enable or disable extended thinking. With adaptive thinking, the model decides when deeper reasoning helps. At the default effort level it uses extended thinking when useful; at lower levels it is more selective.

When it helps: mixed workloads where some requests are trivial and others are hard. You no longer need separate configurations for each.

Effort levels

There are four levels: low, medium, high (default) and max. Anthropic notes the model can overthink simple tasks at the default and recommends trying medium when that happens. See the effort documentation.

A sensible starting point:

Task type	Suggested effort
Quick lookups, formatting, short answers	low
Everyday coding and writing	medium
Debugging, design decisions, unfamiliar code	high
Hardest problems where cost matters less	max

Context compaction (beta)

Long agent sessions eventually fill the context window. Compaction automatically summarises and replaces older context when the conversation approaches a configurable threshold, so the agent can keep working. The trade-off is that summarised details can be lost; keep key requirements in a file or system prompt the agent can re-read.

1M context and 128K output

Opus 4.6 offers a 1M-token context in beta on the Claude Developer Platform, with premium pricing above 200K tokens ($10 input / $37.50 output per million). Output can reach 128K tokens. Anthropic reports 76% on the 8-needle 1M variant of the MRCR v2 retrieval test, versus 18.5% for Sonnet 4.5, suggesting much less “context rot” than earlier models.

Agent teams in Claude Code

Claude Code gained agent teams as a research preview: several agents work in parallel and coordinate, which Anthropic says is best for independent, read-heavy work such as codebase reviews. You can take over any subagent directly. See the Claude Code documentation.

Tips: split work into parts that do not edit the same files; give each agent a clear scope; review the combined result before merging.

Cost control checklist

1. Start at medium effort and raise it only for tasks that fail.
2. Keep prompts under 200K tokens unless the extra context clearly helps, to avoid premium pricing.
3. Use compaction for long sessions, but persist critical instructions outside the conversation.
4. Use agent teams for parallel, read-heavy tasks rather than tightly coupled edits.
5. Check current rates on the pricing page; list price at launch was $5/$25 per million tokens.

Example workflows that suit autonomy

Anthropic’s announcement describes several patterns where longer autonomous runs paid off, based on early-access partners (Anthropic):

• Large migrations. One partner described a multi-million-line codebase migration that the model planned up front, adapted as it learned, and finished in half the expected time.
• Issue triage. Another reported the model closing 13 issues and routing 12 more to the right team members in a single day across six repositories, while escalating to a human when needed.
• Parallel subtasks. Several partners noted the model breaking work into independent subtasks and running tools and subagents in parallel.

What these have in common is a clear goal, a way to check results (tests, issue status, review), and a human who reviews the output. Autonomy works best when success can be verified.

When to keep a human closer

• Irreversible actions, such as deleting data, deploying to production or sending messages externally, should require confirmation.
• Ambiguous requirements are better clarified up front; a long autonomous run on a misunderstood goal wastes tokens.
• Security-sensitive environments should run agents with least-privilege credentials and in sandboxes.

Limits to keep in mind

The 1M context is beta and limited to Anthropic’s own platform. Agent teams are a research preview. And as with any model, autonomous runs should be checked with tests and code review before results are trusted.