DevPass Code 2.0 ships a Smart Router: a decision model picks the Claude tier for every message
LLM Gateway released DevPass Code 2.0 on 15 October 2026, built on opencode v2. Its default model is a Smart Router: before each message, a decision model reads the request and answers one choice question, which tier should handle it. The tiers default to claude-haiku-5-5 (fast), claude-sonnet-5-5 (balanced) and claude-opus-5-5 (frontier), and can be changed with DEVPASS_SMART_ROUTER_MODELS.
Which decision models. DevPass Code offers every model LLM Gateway serves on /v1/decisions: jev-1.13.0 (the default, $0.042 per million input tokens, 64K context), microsoft-decision-1 ($0.042, 32K) and gpt-6-luna-decisions ($0.10, 1.05M). /smart-router or ctrl+t switches between them. A decision reads at most the last 8,000 characters of the message, which LLM Gateway puts under $0.0001 on Jev or Microsoft-Decision-1.
Behaviour. One decision per message, so tool calls in a turn stay on the same model and the provider’s prompt cache keeps working. Session titles and context compaction always use the fast tier. If the decision call fails, the turn falls back to the balanced tier. Choosing any other model turns the router off.
What it is, and isn’t. This is a production use of decision models for routing, which is the use case the catalog tracks, but it is not a new model, and the post gives no accuracy or cost-saving numbers for the router itself. The three routing examples come from the vendor’s own demo logs. LLM Gateway’s separate smart model string, in beta and not yet on DevPass, also uses a Jev classifier (see its routing benchmark). No scores changed.