v1.4.0 — Haiku 4.5 model update

What's Changed

• Updated LIGHT tier from claude-3-5-haiku-20241022 to claude-4-5-haiku-20241022
• Haiku 3.5 was removed from the Anthropic API, causing 404 errors on all LIGHT-tier routed requests

Full Changelog: v1.3.0...v1.4.0

v1.3.0 — Auto-strip thinking for Haiku

Fixed

• Strip thinking params when routing to LIGHT tier (Haiku 3.5)
• Prevents 400 errors when callers have thinking enabled globally

Why

Haiku 3.5 doesn't support thinking/extended-thinking. The router now transparently strips these params for Haiku while preserving them for Sonnet/Opus.

v1.2.0

What's New

Model Upgrades

Default tiers now use the latest Claude model family:

• MEDIUM: claude-sonnet-4-20250514 → claude-sonnet-4-6
• HEAVY: claude-opus-4-20250514 → claude-opus-4-6
• LIGHT: claude-3-5-haiku-20241022 (unchanged)

Cost entries added for new model IDs. Previous model IDs retained in costs for backward compatibility.

Update Guide in README

New Section 5: Updating covers three ways to update your installation:

• Ask your OpenClaw agent
• git pull + restart
• Re-run the install script

Also includes guidance on handling custom config.json changes (stash or use ROUTER_CONFIG env var) and how to check for available updates.

README Refresh

• Centered logo with badges
• Renamed to OpenClaw Router
• ibl.ai logo hyperlinked

Upgrading

If you're already running v1.1.0:

1. git pull in your router directory
2. Restart the service: sudo systemctl restart iblai-router
3. Verify: curl http://127.0.0.1:8402/health — should show the new model IDs

Or re-run the install script:

bash scripts/install.sh

Full Changelog: v1.1.0...v1.2.0

v1.1.0

What's New

Scoring Fix: Skip System Prompt

OpenClaw sends a large, keyword-rich system prompt with every request (tools, skills, memory, identity, etc.). Previously, extractText() included body.system, causing every request to score ~0.44 and route to Opus regardless of actual task complexity.

Now the scorer only evaluates user messages. Routine tasks correctly score 0.01–0.20 and route to Sonnet. Complex tasks still route to Opus.

Install/Uninstall Improvements

• Install script auto-patches agents.defaults.models allowlist in openclaw.json (fixes model not allowed: iblai-router/auto)
• Uninstall script fully cleans up: provider config, model allowlist, cached models.json, legacy smart-router entries

Documentation

• 30-day savings projection table with real-world numbers
• Troubleshooting guide for common issues
• How to check your cost savings anytime

Upgrade

cd router && git pull
sudo systemctl restart iblai-router

Numbers

~80% cost savings on cron-heavy workloads (ops monitoring, inbox processing). 173 requests: $3.97 vs $19.42 baseline.