Anthropic releases Claude 4.5 Agent Mode: native code execution and 30% price reduction

Core conclusion

Anthropic officially released Claude 4.5 Agent Mode today, adding native Agent capabilities to the Claude 4 series: code execution, web browsing, file reading and writing, and multi-step autonomous tool call chains (10+ consecutive tool calls). At the same time, the API price is reduced by 30%, and the new price is $8/1M input Token and $25/1M output Token.

Key Points

• Published: May 18, 2026 -Affected objects: AI automation developers, content production teams, SaaS entrepreneurs
• Core change: Claude is upgraded from a dialogue model to an Agent platform that can perform tasks independently

Background and trigger events

In the past year, competition on the AI Agent track has become increasingly fierce. OpenAI's GPT-5 has long supported code execution and tool invocation. DeepSeek V4 has just released a 1.5T parameter model under the MIT open source license. Google Gemini 3 has also launched a 10M context window. As a safety-first AI company, Anthropic has been relatively conservative in its Agent capabilities. The release of Claude 4.5 Agent Mode marks Anthropic’s official entry into the AI ​​Agent battlefield.

Related sources: Anthropic official blog, Ars Technica, HN discussion

Key Impact (by Dimension)

Adaptation suggestions

Act now

• Apply for Claude 4.5 API access and test the actual performance of Agent Mode
• Upgrade existing Claude Code automation scripts to a version that supports Agent Mode
• Evaluate whether $8/$25 pricing is suitable for your use case compared to GPT-5’s $10/$30

Medium to long term

• Consider using Claude 4.5 Agent Mode to replace traditional RPA tools to complete browser automation tasks
• Superimpose Claude 4.5 as the "execution layer" agent in a multi-Agent system
• Track performance changes of SWE-bench and other benchmarks as a basis for model selection

Technical details analysis

The core capabilities of Claude 4.5 Agent Mode include:

1. Native Python execution: Run Python code directly in a sandbox environment, supporting third-party library installation
2. Web browsing: Open web pages independently, extract content, and fill in forms
3. File system operation: read and write local files, support multi-file project management
4. Tool Orchestration: Agent automatically determines the calling order without manual specification.

The SWE-bench score is 72.4%, which is very close to GPT-5's 74.1%, but has a clear advantage in tool calling efficiency (average 5.2 calls per single task vs. 7.8 times for GPT-5).

Tool entry

Claude 4.5 Agent Mode directly affects the use of multiple tools in the current AI Agent ecosystem. Tool keywords such as Claude, Anthropic, OpenAI, GPT-5, DeepSeek, n8n, and Gemini have naturally appeared in the text, and the platform side will match the maintained tools library.

Internal link guidance

• Want to get started? Watch the tutorial: Claude Code automated writing practice: build an AI content production pipeline in 30 minutes
• Real case: He Built an AI Automation Stack with Claude + n8n — $4K to $12K/mo in 6 Months
• Entrepreneurship case: Claude Code 48 hours to start a business: one person + US$29 monthly fee, monthly income in 3 months $9,000