OpenClaw 2026.4.5 Drops With Video, Music, 12 New Languages — Days After Anthropic Cuts It Off
OpenClaw 2026.4.5 drops video, music, /dreaming, and 12 new languages — days after Anthropic cut off Claude subscription access.
OpenClaw just shipped its biggest release yet, and it’s got everything: video generation, music generation, a new /dreaming memory system, and UI support for 12 languages. The drop comes just days after Anthropic effectively banned the agent from using Claude subscriptions — and the timing is hard to ignore.
Creator Peter Steinberger (@steipete) posted the v2026.4.5 release on GitHub on April 6, citing 29 commits and 104 contributors. It’s a reminder that even as Anthropic closes doors, OpenClaw keeps shipping.
AI Goes Multimodal — And OpenClaw Goes First
The headline features in this release are video and music generation, both baked in as native tools. Agents can now spin up video clips via configured providers and create music through bundled Google Lyria and MiniMax integrations. ComfyUI support is also bundled in, giving OpenClaw a proper image, video, and audio pipeline without external scaffolding.
Then there’s /dreaming. It’s an experimental weighted recall system that promotes short-term memories over time through light, deep, and REM phases — like giving OpenClaw something close to actual memory consolidation. Dreams content writes to a top-level dreams.md file instead of daily notes, and there’s a full Dreams UI with a Dream Diary surface. Weird? Yes. Fascinating? Also yes.
Structured task progress — the kind that actually shows agents thinking through steps rather than hallucinating a result — gets its first real outing here too. Prompt caching got a serious overhaul: deterministic MCP tool ordering, compaction, embedded image history, and normalized system-prompt fingerprints so models don’t re-process the same system context on every call.
Anthropic Cut Them Off. OpenClaw Moved On.
On April 4, Anthropic flipped the switch. Claude Pro and Max subscribers could no longer use their flat-rate plans to power third-party agents like OpenClaw. The reason: third-party services don’t optimize for prompt cache hit rates, so Anthropic called it unsustainable. Users got a one-time credit and a 30% discount nudge toward pay-as-you-go billing.
Steinberger was unimpressed. “Funny how timings match up,” he posted on X. “First they copy some popular features into their closed harness, then they lock out open source.” He’s not wrong. Anthropic had just shipped Discord and Telegram messaging directly into Claude Code — channels that made OpenClaw distinctive — before slamming the door on third-party access.
Meanwhile, GPT-5.4 got genuinely better while Anthropic was busy with policy pivots. OpenClaw’s new providers — Qwen, Fireworks AI, StepFun, MiniMax, Amazon Bedrock Mantle, and a fresh OpenAI Codex integration with GPT-5.4-mini — mean the platform isn’t hurting for capable models. OpenClaw didn’t die when Anthropic cut it off. It just kept building.
