OpenClaw's new version upgrades the default Claude model to Opus 4.7, while the Google plugin adds Gemini voice synthesis.

According to Dynamic Watch Beating, the open-source AI Agent framework OpenClaw has released v2026.4.15. The two significant changes worth noting are: updating the default model on the Anthropic side to yesterday's release of Claude Opus 4.7, and adding Gemini text-to-speech capability to the Google plugin.

One noteworthy fix related to security is: for local MEDIA, the trust in tool results was previously only based on the tool name. Now, if a client declares a tool definition with the same name as a built-in tool, it can opportunistically inherit trust for local media. In v2026.4.15, the trust is strictly tied to the original name of a built-in tool registered in the current run and rejects any definitions with names matching built-in tools or those of other client requests. Both HTTP and SSE paths will return `400 invalid_request_error`.

Other fixes include: cleaning up expired dist chunks during CLI upgrades, automatically correcting legacy openai-codex entries to the correct Codex transfer path, restoring inbound image attachment downloads for the BlueBubbles plugin on Node 22+, and decreasing the default startup and skills prompt budgets to reduce long-session context consumption.