Claude 4.5 Opus excels in code generation but does not stop there: computer use, autonomy and efficiency take a leap forward.
If the race for AI were a hellish wheel, today we would be in the Anthropic box. After the explosive release of Gemini 3 Pro last week, this Monday, November 24, it is the turn of the San Francisco start-up to announce the arrival of Claude 4.5 Opus. Anthropic’s new flagship now exceeds OpenAI, DeepMind and xAI on code benchmarks.
Claude 4.5 Opus, a “senior engineer”
After Haiku 4.5 (light version) and Sonnet 4.5 (intermediate version), Opus 4.5 stands out as the heaviest and most efficient version of the Claude range. The model clearly raises the level on four axes: code, agentics, advanced office automation and computer use. According to Anthropic, Opus 4.5 now approaches programming problems like “a senior engineer”, that is, capable of handling ambiguity, reasoning across multiple systems and finding the necessary fixes alone.
In terms of benchmarks, Claude 4.5 Opus clearly sits where Anthropic wants to concentrate its verticality: code and agentics. On SWE-bench Verified, Claude 4.5 Opus e at 80.9%, ahead of Gemini 3 Pro (76.2%) and GPT-5.1 (77.9%). On terminal coding (ability to act on a terminal), it maintains the advantage with 59.3% against 54.2% for Gemini. On agentic tool use (i.e. agentic tool calls), Opus reached 98.2%, the best performance.
On the other hand, on raw reasoning or multimodal tasks, Gemini remains slightly ahead: Opus scores at 87.0% in reasoning when Gemini exceeds 91%, and at 60.7% on MMMU vision.
An even more autonomous model
Beyond raw performance, Opus 4.5 above all marks a leap in agentic autonomy. The model no longer just generates plans: it executes them and handles any errors without any problem. In multi-step use cases, it maintains the context, orchestrates sub-agents and selects the right tools alone, where Sonnet 4.5 still had to be guided. Opus 4.5 would thus be a very good sub-agent orchestrator. The other innovation comes from efficiency. Opus 4.5 achieves its results by consuming significantly fewer tokens than previous versions. At equal effort, it outperforms Sonnet by using up to 76% fewer tokens.
In terms of prices, Claude 4.5 Opus costs 5 dollars for input and 25 dollars for output per million tokens, which is much more than a Gemini 3 Pro at 2 dollars for input and 12 for output, and even significantly above GPT-5 ($1.25/$10). Pricing which therefore remains attractive if we are to believe the optimization of Claude 4.5 Opus.
Changes in the Claude ecosystem
In its logic of increasing power in agentics, Anthropic is also pushing its product integrations. Claude for Chrome, until now reserved for a few testers, becomes accessible to all Max subscribers ($100/month). The extension allows Opus 4.5 to act directly in the browser, opening, reading, manipulating and crossing tabs to perform complex tasks.
Finally, Anthropic also drastically improves one of the main friction points for chatbots: context management. In the Claude application, Opus 4.5 now automatically summarizes the context when necessary. Very concretely, users will be able to continue their exchanges without loss of information. The principle is promising: almost unlimited continuity in discussions. But Anthropic remains discreet about the internal mechanics.
As of November 24, Anthropic is deploying its new model in the Claude application, via the Anthropic API, as well as on the three major cloud platforms (AWS, GCP, Azure). If it does not outperform Gemini in certain fundamental verticals (pure reasoning, multimodal vision, etc.), it does, however, consolidate its status as an essential reference for code.
