Anthropic Releases Claude 5 Models with Enhanced Tool Integration
Anthropic has launched version 0.26 of its LLM, introducing three new Claude 5 models: claude-fable-5, claude-sonnet-5, and claude-opus-5. This update leverages LLM version 0.32 to enable new features, including server-side tools for WebSearch, WebFetch, CodeExecution, and AnthropicMCP. These tools are accessible via the LLM's -T interface or Python tools, replacing the older -o web_search* options. The upgrade to llm>=0.32 means that reasoning, tool calls, and tool results now stream as typed events. Reasoning for LLM CLI prompts will be displayed to standard error by default, unless the --hide-reasoning or -R flag is used. The concept of extended thinking has been simplified into 'thinking' and 'thinking_effort', with options ranging from low to max. Claude 5 models engage in thinking by default; however, the -o thinking 0 option can disable this for Sonnet 5 and Opus 5, while Fable 5 always utilizes thinking. The -R/--hide-reasoning flag now ensures that reasoning is omitted from both responses and logs. Previously existing options such as thinking_budget, thinking_display, and thinking_adaptive have been removed in this release.
This update signifies a strategic move by Anthropic to deepen the integration of its large language models with external functionalities. By standardizing server-side tool access through a unified interface, the company is likely aiming to enhance model utility and developer experience, potentially reducing friction in complex task execution. The shift towards streaming typed events for reasoning and tool results suggests a focus on real-time feedback and transparency in model operations. The refinement of 'thinking' parameters indicates an effort to provide more granular control over computational resources and model behavior, balancing performance with efficiency. Over the next decade, such advancements will be critical for LLMs to move beyond conversational agents towards more robust, autonomous systems capable of interacting with the digital and physical world.
AI-generated to prompt reflection — not editorial opinion, not advice, not a statement of fact. How this works.
