Claude Opus 4.5

Opus 4.5 is the release where frontier-tier Claude stopped being a luxury line item. On November 24, 2025, Anthropic reset the price of its top model and tightened how much it wrote to get the same job done. For anyone running Claude Code against real work all day, the math flipped overnight.

Key Specs

What Opus 4.5 Brought to the Table

Three things made the release matter: the price drop, tighter outputs, and training focused on delegation.

A two-thirds price cut. Opus 4 and Opus 4.1 had been running at $15 input and $75 output per million tokens. Opus 4.5 came in at $5/$25. For teams billing API usage at production scale, that was a step change, not a trim.

76% fewer output tokens. Same quality work landed in a fraction of the output. On a task where Sonnet 4.5 might spend 500 tokens, Opus 4.5 often wrapped in about 120. Tighter prose. Sharper edits.

50% fewer tool calls. Each task finished with half the tool round trips of the previous generation. Fewer trips, faster completions, and lower cumulative bills.

Tool Search stacks on top. With Tool Search on, the compounding effect pushed token reduction to around 85%. The model picked the relevant tools up front and skipped the wasted calls to irrelevant ones.

Trained for Sub-Agent Delegation

Anthropic trained delegation into Opus 4.5 on purpose. Writing sub-agent prompts, splitting tasks across workers, and stitching parallel results back together were treated as a direct training objective, not a by-product of general capability.

What you get in practice: cleaner prompts for sub-agents, smarter task distribution, and less information loss when rolling parallel outputs back into one answer. Multi-agent orchestration in Claude Code finally felt dependable with Opus 4.5. Before it, the pattern was fragile.

GitHub noted that Opus 4.5 "surpasses internal coding benchmarks while cutting token usage in half." Replit reported similar numbers on their own evals.

New Capabilities

Effort parameter. A per-request API knob for thinking depth. Low effort on quick work, high effort on complex reasoning. Budget moves without a model switch.

Auto-compaction. At 95% of the context window, older messages get summarized and folded down while the conversation thread stays intact. Long engineering runs stopped hitting a hard ceiling.

1M token context window. Available behind a beta header over the API, the same extended window Sonnet 4.5 had introduced earlier.

How It Compared to Sonnet 4.5

Opus 4.5 was the first Opus release where cost-per-task approached Sonnet. Per-token rates were still higher ($5/$25 against $3/$15), but the 76% drop in output often flipped the total bill in Opus's favor on completed tasks.

On deep reasoning work, the call was easy: Opus 4.5 landed better answers at similar or lower total cost. On the routine end of the job list, Sonnet 4.5 and Haiku 4.5 stayed the cheaper picks.

The Detailed Guide

This page is the model profile and the specs. For a full walkthrough of configuring and using Opus 4.5 in Claude Code, covering the effort parameter, auto-compaction, and multi-agent patterns, see the Claude Opus 4.5 usage guide.

Current Status

Opus 4.5 is still active on the API. Opus 4.6 is the current pick for the newest features and capability gains, but 4.5 keeps working as a reliable, well-understood option for production workflows.

Related Pages

• All Claude Models for the complete model timeline
• Haiku 4.5 for the budget option in the same family
• Opus 4.6 for the next generation
• Opus 4.5 usage guide for hands-on configuration
• Model selection guide for strategic model switching