Claude Opus 5 Brings 1M-Token Context and Five-Level Effort Settings to Agentic Coding

What Shipped

Anthropic’s release notes for Claude Opus 5 confirm a model built around sustained, multi-step work rather than single-turn chat quality. The headline specification is a 1 million token context window — available by default, not as a special add-on — paired with up to 128,000 output tokens per response. For teams working against large codebases or long research threads, that context ceiling determines how much of a repository, ticket history, or document set the model can hold in view at once without external retrieval tricks.

Thinking is on by default in Opus 5, with adaptive thinking that scales reasoning depth to the task. Anthropic also introduced a notable constraint: disabling thinking entirely is only permitted at effort level high or below. Attempting to disable thinking at xhigh or max effort returns an error, signaling that Anthropic now treats extended reasoning as inseparable from its most capable effort tiers.

Pricing held steady at $5 per million input tokens and $25 per million output tokens — the same rate as the prior Opus 4.8 release — meaning the context and capability increase did not come with an on-paper price increase, though larger context windows mean larger bills in practice for context-heavy workloads.

The Five-Level Effort Ladder

The defining interface change in Opus 5 is the effort parameter, which Anthropic now describes as “the primary control” for steering the model’s behavior. The five levels — low, medium, high, xhigh, and max — let developers explicitly choose between fast, cheap responses and slow, exhaustive ones, rather than relying solely on prompt engineering to imply the desired depth.

max effort is positioned for “capability-critical work” — the kind of task where getting the right answer matters more than response time or token cost. This maps directly onto agentic coding workflows: a quick lint-fix suggestion can run at low, while a multi-file refactor across a large codebase, or a security-sensitive migration, can be dialed up to max.

The model also supports task budgets, mid-conversation tool changes, the computer use tool, browser use tool, and an advisor tool — a toolset aimed squarely at agents that operate over extended sessions rather than answering isolated questions.

Why Long-Horizon Context Matters for Agentic Coding

Context-window ceilings have been one of the practical limits on agentic coding tools: an agent working through a large repository, a long chain of test failures, or a multi-day research task eventually runs out of room to hold the full state of the problem, forcing summarization, truncation, or external memory systems that introduce their own failure modes.

A 1M-token window changes the calculus for what fits natively. A window of that size can, depending on tokenization, hold a substantial mid-sized codebase, a long design document plus its revision history, or dozens of related support tickets — all in a single context, without the agent needing to reconstruct earlier state from a compressed summary. Combined with the effort ladder, teams get a lever to control cost for routine tasks while reserving the full context and max reasoning depth for the sessions that actually need it.

Opus 5 is available across the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry, putting the same specification in front of enterprise customers regardless of their existing cloud commitments.

What It Means for Teams Adopting Agentic Coding Tools

The practical shift for engineering teams is less about the headline context number and more about what it removes from the workflow: less manual context curation, fewer dropped threads across long debugging sessions, and an explicit dial (effort level) instead of prompt-engineering workarounds to control how hard the model works on a given task. Teams already running AI coding agents in production should expect the main adoption question to shift from “can it hold enough context” to “how do we budget effort levels across routine versus critical tasks” — a cost and workflow-design problem more than a capability gap.

Frequently Asked Questions

Sources & Further Reading

• Claude Opus 5 Release Notes — Anthropic
• Adaptive Thinking — Anthropic Docs
• Task Budgets — Anthropic Docs
• Computer Use Tool — Anthropic Docs