OpenClaw 2026.5.27 fixes runtime boundaries and cuts cold turns 2.9x

• OpenClaw 2026.5.27 shipped tighter runtime and security boundaries, plus faster gateway and reply paths, according to openclaw’s 2026.5.27 release post and the linked GitHub release.
• The update also broadened provider support: openclaw’s provider thread says OpenAI-compatible embeddings are now core, DeepInfra browsing sees the full credential-aware catalog, and direct Anthropic model IDs resolve; the PixVerse plugin post adds video generation with region selection.
• OpenClaw made its release evidence repo public, and openclaw’s evidence-repo announcement says each release now publishes CI, performance, memory, install, and validation artifacts in a durable ledger.
• Performance work extended beyond one point release: openclaw’s lighter-core summary claims cold agent turns are 2.9x faster, warm turns are 2.5x faster, tarballs are 59% smaller, and dependencies are down 42% from the monthly high.
• The previous day’s 2026.5.26 build already hardened installs and channel flows, with openclaw’s 2026.5.26 post and the follow-up hardening thread calling out Alpine, Windows Scheduled Tasks, Docker timeouts, and npm min-release-age configs.

You can jump straight to the 2026.5.27 release notes, inspect the new release evidence ledger, and browse the live providers docs. The oddest reveal is how much of the speedup story sits outside the headline release post: a separate lighter-core blog post ties the faster turns to a smaller package, fewer dependencies, and more functionality pushed into plugins.

The top-line 2026.5.27 changes were runtime isolation and latency. openclaw’s release post framed them as tighter runtime and security boundaries, faster gateway and reply paths, steadier Codex and app-server memory, plus better channels and provider handling.

That fits the pattern from the 2026.5.26 release, which spent most of its changelog budget on lower-latency replies, more reliable channel flows, and hardened install and Docker paths. In two days, OpenClaw moved the story from reliability cleanup into boundary tightening.

OpenClaw’s provider surface got broader in ways that matter for people wiring mixed-model stacks. openclaw’s provider thread lists five concrete changes:

• OpenAI-compatible embeddings moved into the core provider path.
• DeepInfra browsing now loads the full credential-aware catalog.
• PixVerse added video generation.
• vLLM thinking parameters now work.
• Direct Anthropic model IDs now resolve.

The PixVerse plugin post makes the video path more concrete than the summary tweet does: PixVerse now installs as an OpenClaw video provider plugin, accepts an API key, exposes international and China endpoints, and runs through the shared video tool. The docs linked from the thread live at the providers page and the PixVerse plugin page.

The most useful part of this rollout is probably the public evidence repo. openclaw’s announcement says every release now ships with inspectable CI, performance, memory, install, and validation evidence, rather than a changelog that asks you to trust the summary.

That repo turns release QA into an artifact. steipete’s comment says each release spins up hundreds of CI machines and eventually produces a ledger, and the linked release-evidence page gives OpenClaw users a stable place to inspect what passed.

The day-before release was mostly about making installation and updates less fragile. openclaw’s hardening thread calls out Alpine installs, Windows Scheduled Tasks, npm min-release-age configs, Docker and package timeouts, Parallels update checks, and macOS runner bootstraps.

For users already on the tool, openclaw’s install note kept the update path simple with openclaw update, while fresh installs still use the hosted install script. The interesting part is not the command, it is how much platform-specific failure handling got folded underneath it.

The bigger engineering story sits in package weight and dependency surgery. openclaw’s lighter-core summary claims four measurable deltas:

• Cold agent turns: 2.9x faster.
• Warm turns: 2.5x faster.
• Tarball size: 59% smaller.
• Dependency count: down 42% from the monthly high.

steipete’s thread ties those gains to a leaner dependency stack built around custom pieces for the proxy layer, filesystem safety, image handling in WASM, Opus in WASM, and PDF processing in WASM. An earlier steipete post about libopus-wasm adds one more detail: the meeting-notes and in-meeting voice features rode on replacing older Opus-adjacent dependencies with a WASM implementation, which he said performs roughly on par with native on Node and V8.