OpenClaw version 2026.9.6 is now available. For those unfamiliar with the project: OpenClaw is an open-source AI agent runtime that connects chat platforms, models, tools, memory, skills and other components together. This release brings a massive update: 2614 pull requests, 178 direct commits, from over 350 contributors. The release notes are so extensive that they need to be split into more than a dozen sections.

Let's start with a telling detail. At the top of the release notes page, besides language switching, there are three additional buttons: View as Markdown, Open in ChatGPT, Open in Claude. The project explicitly states that "AI agents and tools can read these release notes as plain Markdown". The changelog is specially formatted and optimized for AI readers. When a project starts writing documentation specifically for agents, it confirms that AI is no longer just end-users — it is now a colleague integrated into the development workflow.

Four new models have been added in this release:

- Claude Opus 5.5. Thinking mode is enabled by default, with 1M token context window and 128K token output. Official catalog pricing is $4 / $20 per million tokens, $0.20 for cache reads, $5 for cache writes.

- GPT-6 Sol and Luna. Two tiers based on the same GPT-6 foundation: Sol costs $2 / $10 per million tokens, Luna costs $0.10 / $0.50 per million tokens — a 20x price difference. Separate pricing applies for inputs over 272K tokens.

- Grok 4.7. Supports both text and image inputs, with reasoning levels ranging from low to extra high. It is the default option for new xAI accounts.

- Xiaomi MiMo V2.6 has also been added to the model catalog, with Pro, Flash, and Pro UltraSpeed variants.

Beyond the new models, several structural updates are even more noteworthy:

Restart resumption. After Gateway restarts, incomplete conversations can resume with their full history, progress and tool outputs retained. The agent autonomously decides how to handle interrupted sub-agents: it first checks whether previous work has truly stopped and whether specific operations took effect, then decides whether to retry or switch to an alternative approach. For agents running long-running tasks, this is a major leap forward in solving the long-standing problem of lost progress after restarts.

Decision Models: a model role independent of chat conversations, disabled by default. TypeSafe Jev provides hosted inference, the ONNX plugin runs inference locally on CPU, and Kev connects to your own self-deployed local service. Plugins can use this for structured selection, information scoring, and condition checking without consuming chat context. Not all model calls need to go through a conversation flow — this architectural approach is quite interesting.

Remote workspaces. Files, Memory, and Skills can now be hosted on a separate machine from the agent's main workspace. Files can be pulled from the remote host to chat, memory files are searched and maintained on the remote, and skills can install dependencies remotely. Gateway and node versions must match, and the default single file size limit is 50 MiB.

GitHub reader pulls public issues, PR discussions, commits and diffs directly alongside your chat, supporting up to ten tabs in read-only mode. Check results are provided as snapshots, and do not constitute a readiness-for-merge judgment.

The Web UI now adds a full 30-day usage report, which breaks down tokens, estimated cost and number of sessions by session starter. Costs are explicitly marked as estimates, not an official billing statement.

There are also a number of tangible improvements for daily use:

- Long chats load faster, reading position stays stable, and streaming responses no longer jump the viewport away from your current reading spot.

- Telegram group history is preserved after /reset, so agents can refer back to earlier discussions.

- Meeting notes update in real time during recording, refreshing approximately every five minutes.

- Voice memos can be played directly within chat.

- The maximum size for HTML attachment preview has been increased to 2 MiB.

- Headless nodes automatically update when idle, checking once per hour with at least 12 hours between two activation attempts.

Two breaking changes to watch out for. Code Mode is now changed to pure JavaScript execution — existing TypeScript units will stop working. To migrate, remove the `language` and `typecheck` parameters. SIGUSR1 is now reserved for the debugger, so the Gateway restart signal has been changed to SIGUSR2. If you have custom automation scripts, remember to update this setting.

One final observation. This changelog includes 1175 entries for maintainer and internal changes, which takes up more space than the feature updates themselves: test infrastructure, CI optimization, fixture isolation, all covered in a full dedicated chapter. For a project that merges thousands of pull requests per day, investment in test reliability and release pipelines is just as important as feature development. This isn't the flashiest part of the release, but it is likely why all 2614 pull requests could be shipped stably.

Full changelog: [OpenClaw v2026.9.6 Release Notes](https://docs.openclaw.ai/releases/2026.9.6)