Anthropic released Claude Opus 4.5 on November 24, 2025, as the new flagship model in the Claude 4.5 family. It is the company’s most capable model for coding, agents and computer use, and it is also better at everyday work such as research, slides and spreadsheets. It scored 80.9 percent on SWE-bench Verified in Anthropic’s published evaluations, the first model to pass 80 percent on that coding test. API pricing dropped to $5 per million input tokens and $25 per million output tokens, down from $15 and $75 for the earlier Opus 4.
What changed in Claude Opus 4.5
Claude Opus 4.5 is Anthropic’s new flagship model, released on November 24, 2025. It leads on coding tests with 80.9 percent on SWE-bench Verified, adds an effort setting to trade speed against depth, and costs $5 per million input tokens and $25 per million output tokens. It is available in the Claude apps, the API and the three cloud platforms.
Opus 4.5 replaces Opus 4.1 as the top model in the Claude lineup. It also handles vision, reasoning and math better than earlier models, and it is stronger at deep research, slides and spreadsheets. The model ID for developers is claude-opus-4-5-20251101, and it accepts text, image and PDF input.
Anthropic also updated the Claude Developer Platform, Claude Code and the consumer apps at the same time. The platform adds tools for longer-running agents, Claude Code gets a better planning mode and a desktop app, and the Claude apps stop cutting off long conversations. The model is stronger, the tools around it run longer tasks with less hand-holding, and the price is lower than the model it replaces.
Opus 4.5 sits above Claude Sonnet 4.5 in the lineup, and we covered Sonnet 4.5 in a separate article. Opus 4.5 costs more per token than Sonnet, but Anthropic says it uses far fewer tokens to reach the same result, so the cost of a finished task can be close. It is available on Amazon Bedrock, Google Cloud Vertex AI and Microsoft Foundry, as well as through Anthropic’s own apps and API.
Coding performance and the SWE-bench score
Claude Opus 4.5 scored 80.9 percent on SWE-bench Verified in Anthropic’s published evaluations, the first model to pass 80 percent on the test. SWE-bench Verified checks whether a model can fix real bugs from open source projects, so the score reflects practical coding skill rather than textbook questions.
SWE-bench Verified is a set of 500 real GitHub issues from open source projects. The model has to read the issue, find the fault, and write a fix that passes the project’s own tests. A score of 80.9 percent means it fixed about four out of five issues on its own. YourStory notes that media coverage also called it the first model to pass 80 percent. The score is an Anthropic-reported result, not an independent test.
Anthropic also tested the model on its own hiring exam for performance engineers, a difficult take-home exam with a two-hour limit. Opus 4.5 scored higher than any human candidate ever had. Anthropic notes that the exam tests technical ability and judgment under time pressure, not collaboration or communication, so the result does not mean the model replaces engineers.
The model also scored 66.3 percent on OSWorld, a test of real-world computer use, according to YourStory. That test asks the model to carry out tasks on a real computer, such as using apps and files, rather than answering questions. For teams that write software every day, the coding score is the number that matters most.
The effort parameter and token efficiency
The new effort parameter lets developers choose how hard the model thinks about a task. At a medium effort level, Opus 4.5 matches Sonnet 4.5’s best SWE-bench Verified score while using 76 percent fewer output tokens. At the highest effort level, it beats Sonnet 4.5 by 4.3 percentage points while using 48 percent fewer tokens.
Some tasks need deep thought, and some need a fast answer. The effort parameter lets the developer set that trade-off on each API call instead of picking a different model. A quick check can run at low effort, and a hard bug can run at high effort. Anthropic says the model uses far fewer tokens than its predecessors because it backtracks less and explores less dead ground.
Opus 4.5 costs $5 per million input tokens and $25 per million output tokens, according to Anthropic’s pricing docs. That is a third of the $15 and $75 that Opus 4.1 cost. The batch API cuts the price in half again, to $2.50 and $12.50 per million tokens. Cached input costs $0.50 per million tokens to read, and $6.25 per million to write.
The savings add up on long jobs. A coding agent that runs for an hour can burn through millions of tokens, so a model that uses 76 percent fewer output tokens at the same quality changes the bill in a real way. Anthropic says the combination of lower prices and fewer tokens makes Opus-level ability open to more users, teams and companies.
Long context, memory and multi-agent work
Claude Opus 4.5 has a 200,000-token context window and can return up to 64,000 tokens in a single response. Anthropic added context compaction, context management and memory tools so the model can run longer jobs without losing track of the task.
A 200,000-token context window holds roughly 400 pages of text, according to the Puter model card, which is enough for a large codebase or a long set of documents in one session. The limit is smaller than the 1 million tokens that some rival models offer. Opus 4.5 makes up for it with better handling of what is inside it. Context compaction squeezes earlier parts of a long job so the model can keep going, and the memory tools let it store facts across sessions.
Anthropic also says Opus 4.5 is effective at managing a team of subagents, which lets developers build multi-agent systems that split a large job into parts. In Anthropic’s testing, the combination of effort control, context compaction, advanced tool use, context management and memory raised the model’s score on a deep research evaluation by almost 15 percentage points.
The model can also show surprising judgment. On a benchmark called tau2-bench, the model had to act as an airline service agent for a distressed customer. The test expected it to refuse a change to a basic economy booking. Instead, Opus 4.5 upgraded the cabin first and then changed the flights, which solved the problem in a legitimate way the test had not expected. The benchmark scored that as a failure, but Anthropic presents it as an example of the model finding a valid path the test did not plan for.
Claude Code and the developer platform
Claude Code gets two upgrades with Opus 4.5. Plan Mode now builds more precise plans and asks clarifying questions first, and Claude Code is now available in the desktop app, where users can run several local and remote sessions in parallel.
Plan Mode is the part of Claude Code that plans before it writes code. It now asks the user questions up front, then builds a plan file the user can edit before the model starts work. The desktop app lets a developer run one agent to fix bugs, another to research GitHub, and a third to update docs, all at the same time.
The Claude Developer Platform adds the effort parameter, context compaction and advanced tool use around the model. Anthropic says developers can pick the building blocks they need and control efficiency, cost and capability for each job. The model is available through the API as claude-opus-4-5-20251101, and through Amazon Bedrock, Google Cloud Vertex AI and Microsoft Foundry.
Opus 4.5 also supports function calling, structured output, image input, PDF input, reasoning, prompt caching, web search and computer use, according to the LiteLLM model page. Those features let the model act as an agent rather than a chatbot. It can call tools, read files, search the web and operate a computer, then report back. For a comparison with the other flagship models, see our OpenAI GPT-5 and Google Gemini 3 articles.
Claude for Chrome and Claude for Excel
Claude for Chrome is now available to all Max users, and Claude for Excel has expanded its beta to all Max, Team and Enterprise users. Both tools let the model work inside the apps people already use, instead of asking them to switch to a chat window.
Claude for Chrome lets Claude handle tasks across browser tabs. The user can ask it to read pages, fill in forms, compare options or carry out multi-step tasks on the web. Anthropic opened it to all Max users with the Opus 4.5 launch. Browser automation needs care, because the model acts on live pages, so users should check what it does before it completes a purchase or changes an account.
Claude for Excel works inside spreadsheets. The user can ask it to build formulas, clean data, create charts or check figures, and it works on the file directly. Anthropic announced it in October 2025 and expanded the beta with the Opus 4.5 release. The tool fits finance and operations teams that live in spreadsheets.
The two tools show where Anthropic is heading. The model is moving out of the chat box and into the software people use for real work. The gains in computer use, shown by the 66.3 percent OSWorld score, are what make that move possible. A model that can use a browser and a spreadsheet reliably is more useful than one that only writes text.
Everyday tasks: research, slides and spreadsheets
Opus 4.5 is also better at everyday work. Anthropic says it is meaningfully better at deep research and at working with slides and spreadsheets, and long conversations in the Claude apps no longer hit a wall because the model now summarizes earlier context on its own.
Deep research is the kind of task where the model reads many sources and writes a report. Opus 4.5 handles ambiguity better than earlier models, according to Anthropic’s testers, and it reasons about trade-offs without being told each step. Testers said tasks that were near-impossible for Sonnet 4.5 a few weeks earlier were now within reach.
In the Claude apps, the model now summarizes earlier context automatically as needed, so the chat can keep going instead of stopping when the context window fills up. YourStory calls this “endless chat” for paid users. That change matters for research and writing jobs that run over days, where the user needs the model to remember what was decided earlier.
The model is also stronger at vision, reasoning and math. The Puter model card lists scores of 81.0 percent on GPQA Diamond for graduate-level science questions and 62.7 percent on AIME 2025 for advanced math, sourced from Artificial Analysis. Those scores are third-party figures, not Anthropic’s own, so they should be read as one outside view rather than an official result.
Safety, alignment and prompt injection
Anthropic says Claude Opus 4.5 is the most robustly aligned model it has released, and it is harder to trick with prompt injection than any other frontier model in the industry. The full capability and safety evaluations are in the Claude Opus 4.5 system card.
Prompt injection is the attack where hidden instructions in a web page, a file or an email try to make the model do something the user did not ask for. It is a real risk for agents that read the web and act on what they find. Anthropic says Opus 4.5 has made substantial progress against these attacks, and that it is harder to trick than any other frontier model. That claim is Anthropic’s own assessment, published in the system card.
The system card also covers how the model handles requests that could cause harm and how it acts when it is unsure. Anthropic says its customers use Claude for critical tasks where they need to trust the model. The safety work matters more as the model takes on longer jobs with more tools, because a model that can act on its own needs stronger guardrails than one that only writes text.
Users should still check the model’s work. A model that is harder to trick is not one that cannot be tricked, and a model that writes good code can still write a wrong fix. The safety gains lower the risk, but they do not remove the need for a human to review important output.
Price, plans and where to use it
Claude Opus 4.5 costs $5 per million input tokens and $25 per million output tokens through the API, with batch requests at half price. It is available in the Claude apps on Pro, Max, Team and Enterprise plans, through the API, and on Amazon Bedrock, Google Cloud Vertex AI and Microsoft Foundry.
The $5 and $25 prices are the standard API rates, confirmed in Anthropic’s pricing docs. They are a large cut from the $15 and $75 that Opus 4.1 cost. The batch API charges $2.50 per million input tokens and $12.50 per million output tokens, for jobs that can wait. Prompt caching costs $0.50 per million tokens to read cached input and $6.25 per million to write it.
In the Claude apps, Anthropic removed the Opus-specific usage caps, and it raised overall usage limits for Max and Team Premium users. Anthropic says those users will have roughly the same number of Opus tokens as they had with Sonnet before. The limits are specific to Opus 4.5, and Anthropic says it will update them as future models pass it.
The model is available on all three major cloud platforms, which matters for companies that already run their work on AWS, Google Cloud or Azure. Developers can call it through the API with the model name claude-opus-4-5-20251101. Users who want to try it without an API key can use the Claude apps. Prices and limits can change, so developers should check Anthropic’s pricing page before they commit a workload to the model.
Frequently Asked Questions
When was Claude Opus 4.5 released?
Anthropic released Claude Opus 4.5 on November 24, 2025. It was available the same day in the Claude apps, through the API, and on Amazon Bedrock, Google Cloud Vertex AI and Microsoft Foundry.
What is the SWE-bench Verified score for Claude Opus 4.5?
Anthropic reports 80.9 percent on SWE-bench Verified, which made Opus 4.5 the first model to pass 80 percent on the test. The test checks whether a model can fix real bugs from open source projects. The score is an Anthropic-reported result, not an independent test.
How much does Claude Opus 4.5 cost?
The API costs $5 per million input tokens and $25 per million output tokens. The batch API costs $2.50 and $12.50 per million tokens. Prompt caching costs $0.50 per million tokens to read and $6.25 per million to write. In the Claude apps, the model is included in the Pro, Max, Team and Enterprise plans.
What is the context window of Claude Opus 4.5?
Claude Opus 4.5 has a 200,000-token context window and can return up to 64,000 tokens in a single response. That is roughly 400 pages of text in one session. Anthropic added context compaction and memory tools so the model can run longer jobs without losing track of the task.
What is the effort parameter in Claude Opus 4.5?
The effort parameter lets developers choose how hard the model thinks about a task. At a medium effort level, Opus 4.5 matches Sonnet 4.5’s best SWE-bench Verified score while using 76 percent fewer output tokens. At the highest effort level, it beats Sonnet 4.5 by 4.3 percentage points while using 48 percent fewer tokens.
How is Claude Opus 4.5 different from Claude Sonnet 4.5?
Opus 4.5 is the flagship model and Sonnet 4.5 is the mid-tier model in the same family. Opus 4.5 scores higher on coding tests, handles longer and harder tasks, and costs more per token. Sonnet 4.5 is cheaper and faster for everyday work.
Is Claude Opus 4.5 available on the cloud platforms?
Yes. Opus 4.5 is available on Amazon Bedrock, Google Cloud Vertex AI and Microsoft Foundry, as well as through Anthropic’s own apps and API. Developers can call it with the model name claude-opus-4-5-20251101.
Sources
- Anthropic. “Introducing Claude Opus 4.5.” November 24, 2025. https://www.anthropic.com/news/claude-opus-4-5
- YourStory. “Anthropic introduces Claude Opus 4.5 with hybrid reasoning, benchmark gains, new Excel/Chrome tools.” https://yourstory.com/ai-story/anthropic-claude-opus-4-5-excel-chrome-tools-ai
- LiteLLM. “Claude Opus 4.5 API Pricing.” https://models.litellm.ai/models/claude-opus-4-5
- Anthropic. “Pricing – Claude Platform Docs.” https://platform.claude.com/docs/en/about-claude/pricing
- Puter Developer. “Claude Opus 4.5 – API, Specs, Playground and Pricing.” https://developer.puter.com/ai/anthropic/claude-opus-4-5/
