Technology & AI
Ilya Berdysh
Dec 19, 2025
·
Updated on
Dec 19, 2025
On September 29, 2025, Anthropic released Claude Sonnet 4.5—model that at release time became the world's best for programming, creating autonomous agents and computer control. This is a balanced model combining high performance with reasonable price.
Sonnet 4.5 showed 77.2% on SWE-bench Verified test and 61.4% on OSWorld (computer control test). Price remained same—$3 per million input tokens and $15 per million output, like Sonnet 4.
What Is Claude Sonnet 4.5
Claude Sonnet 4.5 is a large language model from Anthropic, optimized for complex programming tasks, creating autonomous agents and long-term projects. This is the "workhorse" of Claude lineup—not as powerful as Opus 4.5, but significantly faster and cheaper.
Model capable of maintaining task focus for over 30 hours without quality loss. This is critical for complex multi-step projects—refactoring large codebases, migrations between technologies, and long-term research.
Sonnet 4.5 available through Anthropic API (identifier claude-sonnet-4-5), claude.ai web application, iOS and Android mobile apps, and through cloud platforms Amazon Bedrock, Google Vertex AI and Azure. The context window is 200,000 tokens (beta version with 1,000,000 tokens available), maximum output up to 64,000 tokens, knowledge base current through January 2025.
Main Sonnet 4.5 Achievements
At release time September 29, 2025, Sonnet 4.5 was the world's best model for programming with 77.2% result on SWE-bench Verified. Using high compute (parallel attempts and selecting best solutions), the model achieves 82.0%—this was an absolute record until Opus 4.5 release.
On the OSWorld test, which checks the ability to control real computer tasks, Sonnet 4.5 showed 61.4%. This is a huge leap compared to Sonnet 4, which reached 42.2% just four months ago. Growth was 45% in a short period.
The model also showed significant improvements in specialized areas. Experts in finance, law, medicine and exact sciences noted dramatically better knowledge and reasoning compared to previous models, including Opus 4.1.
Extended Thinking and Long Context
Sonnet 4.5 supports extended thinking—mode when the model "thinks" before answering, improving the quality of solving complex tasks. Two output modes available: summarized (brief thought summary) and interleaved (thoughts mixed with main answer).
For organizations, a beta version with 1 million token context window available—5 times more than the standard 200,000. Such volume is critical for analyzing huge codebases, processing multiple documents simultaneously or working with very long meeting transcripts.
Security and Alignment
Anthropic claims Sonnet 4.5 is the most aligned frontier model at release time. The company significantly reduced problematic behavior: sycophancy, deception, power-seeking and encouraging misconceptions.
The model also received improved protection against prompt injection—attacks where malicious actors inject harmful instructions into normal queries. This is critical for agents working with external data.
Comparison with Competitors
In late September 2025, the large language model market was very competitive. Claude Sonnet 4.5 released alongside other strong models from OpenAI and Google. Each company claimed breakthrough results, but actual metrics differed depending on the test.
Key comparison parameters—performance on standard tests, price, context size, speed and additional capabilities. For autonomous agents, OSWorld results and ability to maintain long-term tasks are especially important.
Feature | Claude Sonnet 4.5 | Claude Opus 4.5 | GPT-5.2 | GPT-5.1 | Gemini 3 Pro |
Release date | Sep 29, 2025 | Nov 24, 2025 | Dec 12, 2025 | Nov 2025 | Nov 2025 |
Developer | Anthropic | Anthropic | OpenAI | OpenAI | Google DeepMind |
Context | 200K / 1 M (beta) | 200K tokens | 128K tokens | ~200K tokens | 2M tokens |
Max output | 64K tokens | 64K tokens | 64K tokens | ~16K tokens | 64K tokens |
Price in/out | $3 / $15 per 1M | $5 / $25 per 1M | $2 / $10 per 1M | $1.25 / $10 per 1M | $2 / $12 per 1M |
SWE-bench Verified | 77.2% (82.0% high compute) | 80.9% 🥇 | ~78% | 77.9% | 76.2% |
OSWorld | 61.4% 🥇 (at release) | Excellent | Average | Good | Good |
Programming | Excellent | Best | Excellent | Excellent | Excellent |
Agents | Excellent | Best | Good | Good | Good |
Mathematics | Excellent | Excellent | Excellent | Excellent | Excellent |
Speed | Fast | Medium | Fast | Medium | Fast |
Extended thinking | ✅ Yes | ✅ Yes | ✅ Yes (o1) | ✅ Yes (o1) | ❌ No |
Memory | ✅ Beta | ✅ Beta | ✅ Yes | ✅ Yes | ❌ No |
Long-term tasks | 30+ hours | 30+ hours | Average | Average | Average |
Multimodal | Text + images | Text + images | Text + images + audio | Text + images + audio | Text + images + video |
Comparison Conclusions:
Claude Sonnet 4.5 offers the best price-performance ratio for programming and creating agents. For $3/$15 you get 77.2% on SWE-bench (82.0% with high compute) and best OSWorld result (61.4%). Ability to work 30+ hours on a single task makes it ideal for long-term projects.
Pricing and Usage
Claude Sonnet 4.5 cost remained the same as Sonnet 4—$3 per million input tokens and $15 per million output tokens. This makes the model accessible for large-scale projects while maintaining high performance.
Additional savings available through prompt caching (up to 90% discount) and batch processing (50% discount). When caching repeated prompt parts, cost can drop to $0.30/$1.50 per million tokens.
Access Through Applications
Sonnet 4.5 available to all paid Claude users: Pro, Max, Team and Enterprise. Unlike Opus 4.5, which has limits for some tiers, Sonnet is available with generous limits for all.
Model works in claude.ai web application, iOS and Android mobile apps, and through cloud platforms Amazon Bedrock, Google Vertex AI and Microsoft Azure.
What's New in Sonnet 4.5
Along with Sonnet 4.5 release, Anthropic released Claude Agent SDK—infrastructure for creating autonomous agents. These are the same tools the company uses to create Claude Code.
SDK solves complex problems: how agents should manage memory in long-term tasks, how to handle permission systems balancing autonomy and user control, and how to coordinate sub-agents working toward a common goal.
Main pattern—Planner → Worker(s) → Evaluator. One agent plans a task, multiple agents execute parts in parallel, one agent checks results and makes decisions about next steps.
Claude Code Updates
Claude Code—autonomous coding agent—received important updates with Sonnet 4.5 release. Main innovation—checkpoints, one of most requested features. Now can save progress and instantly roll back to previous state if something goes wrong.
Terminal interface updated, native VS Code extension added. Now I can work with Claude Code directly from the code editor without switching between applications.
Code Execution and File Creation
In web applications and mobile apps, Claude can now execute code and create files directly in dialogue. Available: spreadsheets (Excel), presentations (PowerPoint) and documents (Word).
This means can ask Claude to analyze data and create spreadsheets with charts, write reports and format it as a document, or prepare a presentation—all without leaving chat.
Claude for Chrome
Claude for Chrome extension, which uses computer use capabilities, became available to Max users who joined waitlist in August. Extension allows Claude to control browsers: navigate sites, fill forms, work with spreadsheets, execute tasks across tabs.
What Sonnet 4.5 Is Best For
Everyday programming—writing new code, refactoring, writing tests, fixing bugs. Sonnet 4.5 shows excellent results at a reasonable price.
Creating autonomous agents for production. Ability to work 30+ hours and coordinate sub-agents makes Sonnet ideal for complex workflows.
Computer control and task automation in browsers through Claude for Chrome. Best OSWorld result (61.4%) at release time.
Financial analysis, legal research, medical consultations. Experts noted dramatic improvements in domain-specific knowledge.
Processing large document volumes using 1M context (beta). Analyzing entire codebases, multiple PDFs, long transcripts.
Conclusion
Claude Sonnet 4.5 is the best price-performance balance for programming and creating agents at the end of 2025. For $3/$15 you get performance that was top-tier at release (77.2% on SWE-bench, 82.0% with high compute), and ability to work on tasks for 30+ hours.
For most projects, Sonnet 4.5 is preferable to the more expensive Opus 4.5. Performance difference (3-4%) doesn't justify two-fold price increase, unless you need maximum accuracy for mission-critical tasks.
Sonnet 4.5 is a workhorse for developers: fast, reliable, affordable and powerful enough to solve complex tasks.
Frequently Asked Questions (FAQ)
When was Claude Sonnet 4.5 released?
Claude Sonnet 4.5 was released September 29, 2025 (September 30 UTC) by Anthropic. At release time, this was the world's best model for programming and creating agents.
How much does Claude Sonnet 4.5 cost?
Claude Sonnet 4.5 price is $3 per million input tokens and $15 per million output tokens. This is the same price as Sonnet 4. With prompt caching can get up to 90% discount, reducing cost to $0.30/$1.50 per million tokens.
How does Claude Sonnet 4.5 differ from Opus 4.5?
Main differences: Opus 4.5 is more powerful (80.9% vs 77.2% on SWE-bench), has effort parameters and was released two months later. Sonnet 4.5 faster, nearly twice cheaper ($3/$15 vs $5/$25) and available with 1M token context in beta. For most tasks, Sonnet is the best choice.
What is Claude Sonnet 4.5 result on SWE-bench?
Claude Sonnet 4.5 showed 77.2% on SWE-bench Verified in standard mode. Using high compute (parallel attempts and selecting best solutions), the model achieves 82.0%. This was the world's best result until Opus 4.5 released in November.
What is Claude Agent SDK?
Claude Agent SDK is infrastructure from Anthropic for creating autonomous agents. These are the same tools the company uses to create Claude Code. SDK includes solutions for memory management, permission systems and sub-agent coordination.
Is Claude Sonnet 4.5 available in GitHub Copilot?
Yes, Claude Sonnet 4.5 available in GitHub Copilot from October 2, 2025 for all users. The model can be selected in Copilot settings in VS Code, Visual Studio, JetBrains IDEs, Xcode, Eclipse and GitHub Mobile.
What is the context size of Claude Sonnet 4.5?
Standard context window—200,000 tokens, maximum output—64,000 tokens. For organizations in usage tier 4 or with custom rate limits, beta version with 1,000,000 token context available—5 times more than standard.
What is extended thinking in Sonnet 4.5?
Extended thinking is the mode when a model "thinks" before answering, improving quality on complex tasks. Two output modes available: summarized (brief thought summary) and interleaved (thoughts mixed with answer). Can set a token budget for thinking, for example 64K.
Can Sonnet 4.5 work on task for 30 hours?
Yes, Anthropic reports Sonnet 4.5 capable of maintaining focus on complex multi-step tasks for over 30 hours without quality loss. This is critical for refactoring large codebases, migrations and long-term research.
Is Sonnet 4.5 safer than previous models?
Yes, Anthropic claims Sonnet 4.5 is the most aligned frontier model at release time. The company significantly reduced problematic behavior: sycophancy, deception, power-seeking. Also improved protection against prompt injection attacks.
Ilya Berdysh
Dec 19, 2025
