Skip to main content

October 13, 2026 · Brad Gessler

OpenClaw 2026.10.1, GBrain 0.60.175 and Claude Haiku 5.5

Workspaces now run the newest OpenClaw and GBrain, with security fixes in both. Claude Haiku 5.5 joins the model picker at a tenth of Haiku 4.5's price.

We upgraded GBrain workspaces to OpenClaw 2026.10.1 and GBrain 0.60.175, and added Claude Haiku 5.5 to the model picker. The three largest workspaces upgrade this week, in a quiet hour of their own; every other workspace is done.

The versions

PartBeforeNow
OpenClaw2026.9.62026.10.1
GBrain0.60.27.00.60.175.0
Node.js2424
Postgres17.1117.11
pgvector0.8.00.8.0

Claude Haiku 5.5 is in the model picker

Claude Haiku 5.5 replaces Haiku 4.5 among the quick models. It costs $0.10 per million input tokens and $0.50 per million output on prompts up to 100,000 tokens, a tenth of Haiku 4.5. Prompts over 100,000 tokens cost five times that.

Cached input on Claude Sonnet 5.5 now costs $0.10 per million tokens, down from $0.20.

What you might notice

Older memory is labeled “unverified origin”. GBrain now records where each piece of memory came from. Memory saved before the upgrade has no such record, so search shows it as “unverified origin”. Search still finds all of it, and nothing is hidden.

Facts remember who said them. A fact GBrain saves from a conversation now records the person who said it.

Search finds people and companies by their other names. A search for someone also matches the other names GBrain knows them by.

Busy conversations no longer hold up the rest. A long answer in one conversation no longer stalls the others, and a conversation where the assistant ran a command in the background no longer skips the next message.

Scheduled work runs more reliably. A schedule whose model is slow to respond retries instead of skipping that run.

The night librarian runs in every workspace. Workspaces that also chat on Telegram or Slack were missing the nightly pass that turns the day’s long pages into one page per idea. They have it now.

Two more memory tools. GBrain’s MCP server offers 35 tools, up from 33. edit_page replaces exact text in a page instead of resending the whole page.

Security fixes

  • OpenClaw: private context stays out of replies and saved conversations, a turn that saves memory no longer regains unrelated tools, a connection whose credentials can’t be confirmed is refused, file transfers stay inside their folder, and an empty environment variable is no longer used as a credential.
  • GBrain: stored passwords and keys are no longer handed back to an AI reading the memory, private pages and private profile details stay hidden from remote callers, nothing written to memory can fake GBrain’s own safety markers, and a fix to the MCP library stops it sending credentials to an authorization server the MCP server chose.

What you need to do

Nothing, unless you’ve written code that writes to GBrain’s memory through its MCP server or command line. Write tools such as put_page and remember now refuse a parameter they don’t declare, so remove any extra ones. Remote search and query return shorter rows by default; pass fields: "full" to get every field.

The full upstream notes are in the OpenClaw releases and the GBrain changelog.

Don't have a GBrain yet?

Spin one up in a few minutes, and every AI the team uses shares one memory and the same connected accounts.