OpenClaw v2026.2.17 Released: 1M Context Window, Claude Sonnet 4.6 Support, and More

OpenClaw just dropped v2026.2.17, and this one is a big deal. The update lands just days after Peter Steinberger announced he's joining OpenAI — and the community is still shipping fast. Here's everything new in the latest release.

Claude Sonnet 4.6 Is Now Supported

The headline addition is official Anthropic Claude Sonnet 4.6 support (anthropic/claude-sonnet-4-6). OpenClaw now lists it across aliases and defaults, with a forward-compatibility fallback that gracefully handles model catalogs still showing Sonnet 4.5. If your setup hasn't updated upstream yet, OpenClaw will handle the transition without breaking your agent.

1 Million Token Context Window — Opt In Now

Anthropic models now support opt-in 1M context via a new model parameter:

model.params.context1m: true

This maps to Anthropic's anthropic-beta: context-1m-2025-08-07 header and applies to both Opus and Sonnet. If you're running long agentic workflows — think research agents, code review pipelines, or document analysis — this is a massive unlock. More context means fewer truncations and smarter memory across long tasks. For more, see the next major release: v2026.3.8. For more, see our full Claude Sonnet 4.6 review.

New /subagents Spawn Command

You can now launch subagents directly from chat with the new /subagents spawn command. This enables deterministic subagent activation, meaning you don't need to rely on the model figuring out when to delegate — you can trigger it explicitly. Great for power users who build complex multi-agent pipelines.

Smarter Cron Scheduling

The cron system got a major overhaul:

• Deterministic default stagger for recurring top-of-hour schedules (including 6-field seconds cron), so your jobs don't all slam at the same moment
• Auto-migration of existing jobs to persisted schedule.staggerMs
• New flags: openclaw cron add/edit --stagger <duration> and --exact for per-job timing control
• Per-run model/provider usage telemetry logged in cron run logs and webhooks
• A new local usage report script to aggregate token usage by job — finally, proper cost tracking for scheduled agents

Web Tools Get URL Allowlists

You can now set URL allowlists for web_search and web_fetch. This is a security win — you can lock your agent to only fetch from approved domains, reducing prompt injection risk from untrusted web content. Given the security headlines around OpenClaw lately (more on that in our ClawHub malicious plugins article), this kind of control matters. For more, see what changed in the v2026.2.6 security update.

Voice Gets Faster

Voice Call now pre-caches the inbound greeting TTS, so the first playback is noticeably snappier. Minor tweak, but if you're running OpenClaw as a voice assistant, you'll feel it.

Browser and Docker Improvements

• Browser: Now accepts extraArgs config for custom Chrome launch arguments — useful if you need headless browser customization for scraping or automation
• Docker: New optional OPENCLAW_INSTALL_BROWSER build arg to preinstall Chromium and Xvfb right in the image, skipping runtime Playwright installs and speeding up container startup

Bug Fixes Worth Knowing

Two significant fixes in this release:

1. Subagent context overflow — The system now preemptively guards against accumulated tool-result context before model calls. It truncates oversized outputs and compacts the oldest tool-result messages to prevent context-window overflow crashes. If you've ever had a subagent mysteriously fail mid-task, this is probably why.

2. Read tool auto-pagination — The read tool now automatically pages across chunks when no explicit limit is provided, and scales its per-call output budget from the model's context window. Larger context models can now read more before guards kick in.

How to Update

# If you installed via npm
npm update openclaw

# Via Docker, pull the latest image
docker pull openclaw/openclaw:latest

Full release notes are on GitHub.

The Bottom Line

v2026.2.17 is a quality-of-life release with real substance. The 1M context window for Claude and Sonnet 4.6 support alone make it worth updating immediately. The cron telemetry and URL allowlists show the team is thinking about production-grade deployments, not just hobby setups.

With Steinberger now at OpenAI, the open-source maintainers are clearly keeping the momentum going. Update now.