Effort Controls Make Frontier Intelligence a Runtime Dial

2026-06-20·https://www.anthropic.com/news/claude-opus-4-6

articleaianthropicclaudeagent-loopsinference-routingcontext-managementcoding-agents

effort controls plus compaction map directly to joelclaw inference routing, long-running agent workloads, and cost/latency policy

The clever part of Claude Opus 4.6 is not just that Anthropic says the model is better at agentic coding, computer use, tool use, search, and finance work. It is that the release turns how hard should the model think? into an explicit product and API surface.

Anthropic bundles the model with a beta 1M token context window, compaction, adaptive thinking, and effort controls. That is the useful bit for joelclaw: not “always use the biggest model,” but a runtime policy decision about when to spend tokens, when to compact state, when to let an agent think longer, and when to tell it to stop being expensive and get on with it.

The Claude Code agent teams preview is the other signal. Anthropic is productizing parallel subagents, coordination, and takeover via tmux. That sits right next to joelclaw work on agent loops, workflow routing, and inference routing: the model is becoming one actor inside a managed runtime, not the whole damn system.

Key Ideas

Claude Opus 4.6 adds a beta 1M token context window for the Claude Developer Platform, which matters for long-running agents that need more memory without immediate handoff.
Compaction gives Claude a source-level mechanism for summarizing its own context during longer tasks instead of crashing into context limits.
effort controls make intelligence, latency, and cost tunable per request, which maps cleanly to joelclaw model policy and routing decisions.
Adaptive thinking lets the model infer how much extended thinking to use from context, but the release still exposes explicit developer control when “smart” becomes “too slow.”
Claude Code agent teams are positioned for independent, read-heavy work like codebase reviews, which is basically the safe lane for subagent fan-out.
Anthropic reports leading results on Terminal-Bench 2.0, Humanity’s Last Exam, GDPval-AA, and BrowseComp, but the runtime controls are the more durable design pattern than the leaderboard claim.
The Claude Opus 4.6 system card ties capability gains to safety evaluation, including cyber-specific safeguards and probes from next-generation constitutional classifiers.

Links

← All discoveries