# What is Claude Opus 4.5, and what changes with its lower price and effort control?

Date: 2025-11-25 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/what-is-claude-opus-4-5/

**Answer:** Claude Opus 4.5 is the model Anthropic released on November 24, 2025, as its most capable for coding, AI agents and computer use, and also for everyday work with documents, spreadsheets and slides. It is also much cheaper than earlier Opus models, at $5 per million input tokens and $25 per million output tokens, and adds an effort setting that lets developers trade speed and cost against depth.

Yesterday, Monday, November 24, 2025, Anthropic released **Claude Opus 4.5**, the new top model in its Claude family. Claude comes in three sizes: **Haiku**, the smallest and fastest; **Sonnet**, the mid-size model; and **Opus**, the largest and most capable. CNBC notes this is Anthropic's third major model in two months, following Sonnet 4.5 (September) and Haiku 4.5 (October), and that the last Opus, Opus 4.1, came out in August.

Anthropic calls Opus 4.5 its best model for coding, agents and **computer use**, the ability to operate software through the screen as a person would. It also says the model is noticeably better at everyday office tasks: deep research, slides and spreadsheets.

## What Anthropic released

Opus 4.5 is available in the Claude apps, through Anthropic's API for developers, and on the three major cloud platforms. It becomes the default model on the Pro, Max and Enterprise plans, according to CNBC.

The biggest practical change is price. Opus models have been powerful but expensive. The new rates are $5 per million input tokens and $25 per million output tokens; one early customer quoted by Anthropic says earlier Opus models had been too costly to use routinely. A **token** is a small piece of text, roughly part of a word.

On scores, TechCrunch reports that Opus 4.5 is the first model above 80% on **SWE-bench Verified**, a test in which models must fix real bugs from open-source projects. Anthropic also tested it on the take-home exam it gives job candidates for performance engineering. Working within the two-hour limit, the model beat the best result any applicant has ever posted, using a method that runs several attempts and picks the best. Anthropic adds that the exam does not measure skills such as teamwork or communication.

Several product updates came with it:

| Product | Change |
|---|---|
| Claude apps | Long chats no longer stop at the limit; Claude summarizes earlier parts automatically |
| Claude Code | Plan Mode asks questions first, then writes an editable plan.md file; now also in the desktop app |
| Claude for Chrome | Available to all Max subscribers |
| Claude for Excel | Beta opened to Max, Team and Enterprise |
| Usage limits | Opus-specific caps removed; Max and Team Premium limits raised |

## How the effort setting works

A model like Opus can think briefly or at length before it answers. More thinking often gives better results, but costs more tokens and time. Opus 4.5 introduces an **effort parameter** in the API: a dial developers can turn lower or higher to choose how hard the model works on a request.

Anthropic's own measurements show why this matters:

| Effort | Result on SWE-bench Verified, compared with Sonnet 4.5's best score | Output tokens used |
|---|---|---|
| Medium | Same score | 76% fewer |
| Highest | 4.3 points higher | 48% fewer |

In other words, Opus 4.5 often reaches an answer in fewer steps: less backtracking, less repetition. That efficiency is part of why the lower price is possible.

## Longer conversations and longer tasks

The second theme is memory over long work. Every model has a **context window**, the amount of text it can hold in view at once. When a conversation outgrows it, the model has traditionally had to stop.

Opus 4.5 handles this in two ways. In the Claude apps, TechCrunch describes an "endless chat" feature for paid users: when the window fills, the model condenses earlier context and carries on instead of stopping. For developers, Anthropic combines effort control, **context compaction** (the same kind of summarizing) and advanced tool use so agents can run longer with less human help. In an interview with TechCrunch, Dianne Na Penn, who leads research product management at Anthropic, explained that a bigger window is not enough by itself; the model also has to know which details are worth remembering.

Anthropic says Opus 4.5 is also good at managing **subagents**, helper models it hands parts of a job to. TechCrunch describes setups where Opus leads a team of Haiku-powered agents. Here is the kind of request that suits it:

```prompt
Review our 60-page vendor contract and the attached pricing spreadsheet.
List every clause that changes our costs, check each figure against the
spreadsheet, and flag mismatches in a table. Ask me before assuming
anything about renewal dates.
```

> **Try it:** Give Claude a long document you know well, such as a lease or a manual, and ask for a table of key obligations with page references. Check three entries by hand to see how accurate it is.

## Why it matters

Opus 4.5 arrives a week after Google's Gemini 3 and twelve days after OpenAI's GPT-5.1, and CNBC reports Anthropic claims it beats both on SWE-bench Verified. The bigger shift may be price. At the new rates, a top model becomes practical for daily work rather than special occasions.

Anthropic also stresses safety. It calls Opus 4.5 its most robustly aligned model so far and says it is harder to trick with **prompt injection**, hidden instructions in content that try to hijack the model, than any other frontier model on a benchmark built and run by Gray Swan. One example shows the balance involved: in an airline customer-service test, the model found a legitimate workaround, upgrading the ticket's cabin and then changing the flights, that the test itself scored as a failure. Anthropic notes that clever paths around rules can be helpful in one setting and a problem in another.

## When should you choose Claude Opus 4.5, and how do you use its effort setting?

Choose Opus 4.5 for work where quality matters more than speed: complex coding, long documents, financial models, research that combines many sources, or agents that run for a long time. For short, simple questions, a smaller model like Sonnet or Haiku is usually fast enough. In the Claude apps, Opus 4.5 is now the default on paid plans, so you may already be using it.

If you build with the API, start at medium effort. It is cheaper and, by Anthropic's figures, matches Sonnet 4.5's best coding score. Move to high effort only for the hardest tasks, and to low effort for quick, repetitive jobs like sorting or short summaries.

Whatever the setting, give clear goals and limits, ask the model to stop before risky actions, and check its work, especially numbers and code, before you rely on it.

## Sources

- Introducing Claude Opus 4.5: https://www.anthropic.com/news/claude-opus-4-5
- Anthropic releases Opus 4.5 with new Chrome and Excel integrations: https://techcrunch.com/2025/11/24/anthropic-releases-opus-4-5-with-new-chrome-and-excel-integrations/
- Anthropic unveils Claude Opus 4.5, its latest AI model following $350 billion valuation: https://www.cnbc.com/2025/11/24/anthropic-unveils-claude-opus-4point5-its-latest-ai-model.html