Skip to content
AI Primer
release

OpenClaw releases 2026.5.28 with Opus 4.8 support and faster turns

OpenClaw 2026.5.28 added Claude Opus 4.8 and Krea support while cutting fresh-install size 52.8% and speeding both cold and warm turns. It also expanded /subagents inspection, which should make delegated runs easier to debug.

3 min read
OpenClaw releases 2026.5.28 with Opus 4.8 support and faster turns
OpenClaw releases 2026.5.28 with Opus 4.8 support and faster turns

TL;DR

  • OpenClaw 2026.5.28 shipped openclaw's release post with Claude Opus 4.8 support, Krea image generation via fal, faster gateway and session hot paths, and Discord commentary drafts.
  • According to openclaw's benchmark post, version 2026.5.28 made cold turns 14.5% faster, warm turns 16.0% faster, and fresh installs 52.8% smaller than 2026.5.27.
  • The new /subagents view that openclaw's tool note described exposes child runs, labels, sessions, timestamps, and tool logs, which turns delegated-run debugging into a first-class UI surface.
  • OpenClaw's broader May sweep, documented in openclaw's lighter-core summary and linked to the official benchmark post, claims cold agent turns are now 2.9x faster than the monthly baseline, warm turns are 2.5x faster, and tarballs are 59% smaller.

The release notes are all public: the GitHub tag for 2026.5.28 lists the feature adds, the lighter core, sharper claws post publishes the speed and size deltas, and the subagents docs show the new inspection surface. OpenClaw also published a release evidence repo that exposes CI, memory, install, and validation artifacts for each drop.

Models and media

The headline addition is Claude Opus 4.8 support, plus Krea image generation through fal, in the same build that tightened run-time paths. That keeps OpenClaw's model surface moving on both coding and media generation at once.

The previous 2026.5.27 release had already widened provider support. In openclaw's provider roundup, OpenClaw said it moved OpenAI-compatible embeddings into core, added full credential-aware DeepInfra catalog loading, enabled VLLM thinking parameters, and resolved direct Anthropic model IDs.

Turn speed and package size

OpenClaw published a clean before-and-after for 2026.5.28 versus 2026.5.27:

That follows the earlier 2026.5.27 push on gateway, reply-path, and memory stability that openclaw's 2026.5.27 release post summarized as tighter runtime boundaries and steadier Codex and app-server memory.

Subagents

/subagents now exposes the parts of delegated runs that usually vanish into the harness. According to openclaw's tool note, the command shows active child runs, labels, sessions, timestamps, and tool logs.

That is a concrete shift from "subagents exist" to "subagents are inspectable." For teams using OpenClaw as a delegation-heavy coding agent, the interesting part is not the command name, it is that the child-run state and logs are now queryable in one place.

Hot paths

OpenClaw said the 2026.5.28 speedup came from stripping repeated metadata, config, plugin, and session-store work out of active runs. The gateway protocol docs linked from openclaw's hot-path note are the canonical reference for that path.

The same release also made Discord progress drafts show commentary, a small UX change in openclaw's release post that fits the broader theme of making agent state easier to inspect while the run is still live.

Release evidence repo

OpenClaw made its release evidence repository public on May 28. In openclaw's evidence repo announcement, the team said every release now ships with durable CI, performance, memory, install, and validation evidence.

That gives the performance claims a paper trail outside launch tweets. It also means the interesting artifact for future releases may be the evidence bundle itself, not just the changelog blurb or screenshot.

Further reading

Discussion across the web

Where this story is being discussed, in original context.

On X· 3 threads
TL;DR1 post
Models and media1 post
Turn speed and package size2 posts
Share on X