What is Claude Sonnet 4.5?
Claude Sonnet 4.5 is Anthropic's September 29, 2025 Sonnet model, built for coding, agents, computer use, and knowledge work. At release it posted the top SWE-bench Verified score among frontier models and the top OSWorld computer-use score, and Anthropic measured it holding focus on complex multi-step tasks for more than 30 hours. It runs on Unifically as anthropic/claude-sonnet-4-5, accepts text and image input, and outputs text with a 200k context window and 64k max output. It sits before Claude Sonnet 4.6 in the Sonnet family: still a strong daily model for code and agents, now the value pick rather than the front of the line.
Key features of Claude Sonnet 4.5
Leading coding scores at release
It scores 77.2% on SWE-bench Verified, 82.0% with parallel test-time compute, ahead of GPT-5 Codex at 74.5% and Gemini 2.5 Pro at 67.2% at release.
30+ hour task focus
Anthropic measured the model staying on a complex, multi-step task for more than 30 hours. That length changes what one agent run can finish: full features, long migrations, overnight batch work.
Computer use
It scores 61.4% on OSWorld, up from 42.2% for Claude Sonnet 4 four months earlier, and was the best computer-use model of any lab at release.
Broad benchmark coverage
Beyond coding it reaches 50.0% on Terminal-Bench, 98.0% on the τ2-bench telecom split, 83.4% on GPQA Diamond, and 100% on AIME 2025 with Python tools.
Domain knowledge work
It scores 55.3% on the Finance Agent benchmark, ahead of Opus 4.1 at 50.9%, and expert reviews rated its finance, law, medicine, and STEM answers above earlier Claude models.
Alignment gains
Anthropic called it the most aligned frontier model at release, with the lowest misaligned-behavior score in its evaluation and less sycophancy and deception than earlier Claude models.
Best for
Agentic coding
Multi-file fixes, refactors, and feature work that runs for hours without drifting.
Terminal agents
Shell-driven tasks like builds, test loops, and environment setup.
Computer-use agents
Browser and desktop-style tasks that click, type, and read screens, with clear action limits.
Financial analysis
Spreadsheet work, filings review, and research agents, with human review.
Document and chart reading
Image input covers screenshots, PDFs, and figures alongside text.
Cost-aware production agents
Strong results at a mid-line price point when the newest Sonnet is not required.
Use cases
Build a coding agent that takes a ticket, works across a repo, runs the tests, and reports back after hours of unattended work. Use it for customer-support and operations agents that call tools reliably: its 98.0% on the τ2-bench telecom split came from tool-use training that carries over to real workflows. It also fits browser automation like form filling, dashboard checks, and internal-tool operation. For research teams, pair extended thinking with document input to draft financial models, review contracts, or summarize a stack of papers with the reasoning shown.
Limitations
The context window is 200k tokens with 64k max output. Newer Sonnet models carry 1M context and 128k output, so the largest codebases and document sets need chunking here.
Claude Sonnet 4.6 has since passed it on coding, computer use, and long-context reasoning. Sonnet 4.5 remains a capable model, but it is no longer the strongest Sonnet on the same API.
OSWorld's 61.4% still means real UI work fails on a meaningful share of tasks. Production computer-use agents need retries, checkpoints, and action limits.
High-stakes finance, legal, and medical outputs need human review. Expert evaluations rated its domain answers well, and they are still drafts.
Claude Sonnet 4.5 vs Claude Sonnet 4.6
Claude Sonnet 4.6 is the newer model and wins on the numbers: 79.6% versus 77.2% on SWE-bench Verified, and 72.5% versus 61.4% on OSWorld-Verified. It also carries a 1M context window against 200k, 128k output against 64k, and adaptive thinking against Sonnet 4.5's fixed thinking budget. Arena's Text board ranks claude-sonnet-4-6 at #23 with a 1472±4 score, ahead of Sonnet 4.5 at #42 with 1455±3. Pick Sonnet 4.5 when its results already clear your quality bar; pick 4.6 for long context, heavy computer use, and the hardest agent runs.
When to use Claude Sonnet 4.5
Use Claude Sonnet 4.5 for agentic coding, tool-calling workflows, and computer use where a 200k context is enough and you want proven, stable behavior. It is a sensible default for teams already tuned on it. Move to Sonnet 4.6 when a task needs 1M context or the strongest current Sonnet results.
API examples
Call Claude Sonnet 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-sonnet-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-sonnet-4-5",
"messages": [
{ "role": "user", "content": "Refactor this function and write tests that cover the edge cases." }
]
}'
The response comes back synchronously with the completion. Set "stream": true to receive tokens as they generate.
FAQs
People also ask
anthropic/claude-sonnet-4-5, called through the OpenAI-compatible POST /v1/chat/completions endpoint with your Unifically API key.
Text and image input with text output. It reads screenshots, charts, and documents in the same request as code and prose.
200k tokens, with a callable max output of 64k tokens. That fits a mid-size codebase or a long document set in one run, but not the very largest repos.
Yes. It scores 77.2% on SWE-bench Verified and 50.0% on Terminal-Bench, and holds focus on multi-step coding tasks for 30+ hours.
Yes for its generation. It scores 61.4% on OSWorld, the best computer-use result of any model at its September 2025 release. Newer Claude models have since passed it, so set expectations by that baseline.
Yes. Claude Sonnet 4.5 supports extended thinking with a token budget, so you can trade latency for deeper reasoning on hard turns.
The 200k context window is smaller than newer Sonnet models, and finance, legal, and medical outputs still need human review. For the hardest agent work, Sonnet 4.6 is the stronger pick on the same API.
