Claude Opus 4.5
Anthropic's flagship Claude model for coding, agents and computer use
Claude Opus 4.5 is a large language model from Anthropic, released in November 2025 as the company's most capable offering. The company positions it for software engineering, agentic work and computer control, and says it also improved at vision, reasoning and mathematics over earlier Claude models.
- Released November 2025 as Anthropic's top Claude model
- API pricing set at $5 per million input and $25 per million output tokens
- Leads in 7 of 8 languages on SWE-bench Multilingual, per Anthropic
- 10.6% above Sonnet 4.5 on Aider Polyglot, per Anthropic
- Scored above every human candidate on an internal two-hour engineering exam
What it is
Claude Opus 4.5 is a proprietary language model that Anthropic launched in November 2025. It sits at the top of the company's Claude range and is aimed at demanding work: writing and fixing code, running software agents, and operating a computer on a user's behalf. Anthropic also claims gains in research, slide and spreadsheet work, vision, reasoning and mathematics. The company describes it as its strongest model for coding, agents and computer use. It is sold through Anthropic's own apps and API and through the three major cloud platforms, with token pricing set at $5 per million input tokens and $25 per million output tokens.
How it came about
Anthropic announced the model alongside refreshed versions of its developer platform, Claude Code and consumer apps. Before launch, staff testers tried it and, by the company's account, repeatedly said it copes with unclear requirements, weighs trade-offs unprompted and can track down bugs that span several systems. Early customers reported similar results, and several said work that had defeated Sonnet 4.5 shortly beforehand became doable. The company also noted that one internal hiring exam for performance engineers produced a score above any human candidate within the two-hour limit, and said it would publish more on that line of research.
Inside the model
Anthropic has not published the model's size, layer count, training corpus or compute budget in this release. What the company does describe is behaviour: it plans multi-step work, calls tools, and keeps going on long tasks with fewer dead ends. A new "effort" setting lets callers trade thoroughness against speed and cost, and one partner said lower effort still met their quality bar. The company says the model uses fewer tokens than Sonnet 4.5 for the same jobs, and partners reported token savings of up to 65% on held-out tests. It accepts images as well as text and can drive software tools.
What it can do
Anthropic reports leading results on tests of real-world software engineering. On SWE-bench Multilingual it leads in seven of eight programming languages, and it gains 10.6% over Sonnet 4.5 on Aider Polyglot. The company also claims improvements on BrowseComp-Plus for agentic search and 29% higher earnings than Sonnet 4.5 on Vending-Bench, a long-running task test. One partner measured a 15% gain over Sonnet 4.5 on Terminal Bench. In a τ2-bench airline scenario, where the expected answer is to refuse a ticket change, the model instead upgraded the cabin first and then altered the flights, a workaround the company calls legitimate.
How to use it
The model is served through Anthropic's apps, its API and all three major cloud platforms. Developers call it under the identifier claude-opus-4-5-20251101. Token rates are $5 per million in and $25 per million out, which the company frames as bringing Opus-level capability within reach of more teams. No downloadable weights or open licence are offered.
Safety and testing
Anthropic ties the release to research on how AI affects engineering work and other professions, and says findings from its Societal Impacts and Economic Futures group will follow. The company notes that its internal take-home exam measures technical skill and judgement under time pressure only, and does not cover collaboration, communication or experience. Beyond that, the release describes no red-teaming, evaluations or incidents.
Reception and debate
Feedback quoted by Anthropic from testers and early customers is strongly positive. Users praise planning, tool calling, code review and instruction-following, and several report large cuts in token use and in tool-calling and build errors. One firm said accuracy on its Excel and financial modelling evaluations rose 20% with efficiency up 15%. Another said tasks that used to take two hours now take thirty minutes. A partner noted the model scored above every human candidate on a hard two-hour take-home exam, a result the company says raises questions about the future of the profession.
Why it matters
Anthropic presents the release as a step towards AI doing more of everyday knowledge work, and pairs it with product changes: agent tools that run for longer, integrations with Excel, Chrome and desktop apps, and removal of the ceiling on long chats inside the Claude apps. Partners describe it as the first Opus model cheap enough to use as a default, and Notion said it was putting Opus into its agent for that reason. The company frames the model as a preview of broader shifts in how work is carried out.
Questions people ask about Claude Opus 4.5
What is Claude Opus 4.5 for?
Anthropic built it for hard software work: writing and repairing code, running agents that act on their own, and controlling a computer. The company also reports gains in research, spreadsheets, slides, vision, reasoning and mathematics over earlier Claude models.
How do I run it?
Through Anthropic's apps, its API, or the three major cloud platforms. In the API it is called claude-opus-4-5-20251101. Pricing is $5 per million input tokens and $25 per million output tokens. Weights are not downloadable.
How big is it?
Anthropic has not stated a parameter count, layer count, training data size or compute figure for this model in the release, so those details are not available here.
How does it compare with Sonnet 4.5?
Anthropic and its partners report it beating Sonnet 4.5 on coding, agent and long-task tests while spending fewer tokens. One partner measured up to 65% fewer tokens on held-out tests, and another a 15% gain on Terminal Bench.
Is it open source?
No. It is a proprietary model served through Anthropic's apps, API and cloud partners. There are no open weights and no public licence for the model itself.
Good reading and sources
There is no English Wikipedia article on Claude Opus 4.5 yet, so the facts on this page come from the sources listed above: anthropic.com/news/claude-opus-4-5. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.