Skip to content

Claude Opus 4.5 - Model Report

Overview

AttributeDetails
DeveloperAnthropic
Release DateNovember 24, 2025
Model TypeLarge Language Model
LicenseProprietary (Closed Source)

Architecture

SpecificationValue
Total ParametersNot disclosed
Active ParametersNot disclosed
ArchitectureTransformer (details undisclosed)
Context Window200K tokens
Max Output64K tokens
Knowledge CutoffMarch 2025

Training

AspectDetails
Training DataNot disclosed
Training MethodNot disclosed
Safety ApproachConstitutional AI (RLHF)

Key Features

  • Effort Parameter (Beta): Control computational effort allocation across thinking, tool calls, and responses
  • Extended Thinking: Preserves all reasoning blocks across multi-turn conversations
  • Computer Use: Enhanced with zoom action for detailed screen region inspection
  • Agentic Capabilities: Long-running processes, maintains context over hours/days
  • Tool Use: Reliable interaction with external tools and APIs
  • Enterprise Automation: Designed for complex, multi-step workflows

Benchmarks

BenchmarkScore
SWE-bench Verified80.9%
OSWorld (Computer Use)66.3%
MMMU (Visual)80.7%
MMLU90.8%

Pricing

MetricCost
Input Tokens$5 / million
Output Tokens$25 / million
Prompt CachingUp to 90% savings
Batch Processing50% savings

Note: This is 1/3 the price of previous Opus-class models.

Open Source Availability

PlatformStatus
Hugging FaceNot Available
Weights DownloadNot Available
Self-HostingNot Possible

Minimum Hardware for Self-Hosting

RequirementDetails
Self-HostingNot Possible

Claude Opus 4.5 is a closed-source, proprietary model. Model weights are not publicly released and cannot be downloaded or self-hosted.

Access Methods

PlatformAvailability
Claude APIDirect API access
Claude Web/MobileConsumer interface
Amazon BedrockAWS integration
Google Cloud Vertex AIGCP integration
Microsoft Azure FoundryAzure integration

Minimum Apple Product for Self-Hosting

RequirementDetails
Apple ProductN/A
ReasonClosed-source model, weights not available
AlternativeUse via API ($5/$25 per million tokens)

Sources

Technical research and documentation