Claude Opus 4.5

What is Claude Opus 4.5?#

Claude Opus 4.5 is Anthropic's November 24, 2025 Opus model for coding, agents, computer use, and deep reasoning. It closed out the Claude 4.5 family after Sonnet 4.5 and Haiku 4.5, and runs on Unifically as anthropic/claude-opus-4-5. It accepts text and image input and outputs text, with a 200k token context window and up to 64k output tokens. Two things defined this release: it was the first frontier model past 80% on SWE-bench Verified, and it introduced the effort parameter, which lets one model trade speed and token spend against reasoning depth. It also landed at roughly a third of the per-token price of the Opus model before it.

Key features of Claude Opus 4.5#

80.9% SWE-bench Verified

The top agentic coding score at release, ahead of Sonnet 4.5 at 77.2%, Gemini 3 Pro at 76.2%, and GPT-5.1-Codex-Max at 77.9%. It also leads 7 of 8 languages on SWE-bench Multilingual.

Effort parameter

Set low, medium, or high to control how much work the model spends per response. Medium effort matches the best Sonnet 4.5 SWE-bench Verified score with 76% fewer output tokens; high effort beats it by 4.3 points with 48% fewer.

Frontier tool use

It scores 98.2% on tau2-bench Telecom, 88.9% on Retail, and 62.3% on MCP Atlas, a test of scaled tool use where Sonnet 4.5 reaches 43.8%.

Computer use

It reaches 66.3% on OSWorld, up from 61.4% for Sonnet 4.5 and 44.4% for Opus 4.1, and holds up on browser and desktop-style tasks that click, type, and read screens.

Long-horizon coherence

On Vending-Bench, a simulated year of running a small business, it ends with $4,967.06 in net worth versus $3,849.74 for Sonnet 4.5, holding a goal across thousands of steps.

Better prompt-injection resistance

At release it was the hardest frontier model to trick with prompt injection in agentic settings. That matters for agents that read untrusted pages, files, or tool output.

Best for#

Use cases#

Build a coding agent that takes a failing test, traces the bug across a repo, and lands a fix that survives review. Run the same model at low effort for triage and high for the final pass. It also fits research agents that search, read, and reduce dozens of sources into a cited brief, and operations agents that fill forms, reconcile records, or drive internal tools through a browser. For document work, pair its vision input with the 200k window to reason over contracts, filings, and charts together.

Limitations#

The 200k context window is the main constraint today. Newer Claude models hold 1M tokens, so very large codebases or document sets need chunking here that they would not need elsewhere.

Max output is 64k tokens. Long reports may need to be split across turns.

OSWorld is a controlled benchmark. Real web apps are messier, and computer-use agents still need tight action limits and review on anything that changes state.

Prompt-injection resistance improved over earlier models but attacks still land sometimes. High-stakes finance, legal, medical, and security outputs need human review.

Claude Opus 4.5 vs Claude Opus 4.6#

Claude Opus 4.6 is the better choice for long-context and long-running work. It raises the context window from 200k to 1M tokens, output from 64k to 128k, and adds adaptive thinking plus a max effort level above Opus 4.5's high. It also improves Terminal-Bench 2.0 from 59.3% to 65.4%, with gains on ARC-AGI-2, OSWorld, and BrowseComp.

SWE-bench Verified is roughly flat between the two, so Opus 4.5 remains a capable coding model when the task fits in 200k tokens. Pick 4.5 for contained coding and agent work; move to 4.6 when the job is bigger than the window.

When to use Claude Opus 4.5#

Use Claude Opus 4.5 when you want strong agentic coding, tool use, and computer use on tasks that fit within 200k tokens, with effort as a per-request cost dial. Pick a newer Opus model when the job needs 1M context, 128k output, or the deepest reasoning settings. Pick a Sonnet or Haiku model when the task is simple and latency matters more than ceiling.

API examples#

Call Claude Opus 4.5 from any language by POSTing to /v1/chat/completions, the OpenAI-compatible endpoint shared by every language model on the platform. Full parameter docs live at docs.unifically.com/models/llm/anthropic/claude-opus-4-5.

curl -X POST https://api.unifically.com/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -d '{
    "model": "anthropic/claude-opus-4-5",
    "messages": [
      { "role": "user", "content": "Find the race condition in this worker pool and write a failing test for it." }
    ]
  }'

The response comes back synchronously with the completion. Set "stream": true to receive tokens as they generate.