[VickyXAI]

Live bug

SmartChat routed to ids the gateway no longer serves (checked against GET https://blockrun.ai/api/v1/models and a $0 probe of each free id on 2026-10-20):

• Free, Simple/Medium/Complex → nvidia/nemotron-3-super-120b: not in /v1/models. The gateway answers it with nvidia/llama-3.2-11b-vision, so SmartChatResponse.Model named a model that never served the call.
• Free, Reasoning → nvidia/nemotron-3-nano-omni-30b-a3b-reasoning: listed with available: false and also redirected to llama-3.2-11b-vision.
• Eco + Auto, Simple → moonshot/kimi-k2.7: hidden from /v1/models.

Replacements

Profile / tier	Old	New	Why
free / all four	nemotron-3-super-120b, nemotron-3-nano-omni-30b-a3b-reasoning	nvidia/llama-3.2-11b-vision	Only model in the catalog with billing_mode: free, pricing 0 and available: true. The other free ids (nemotron-3.5-lightning, nemotron-3-ultra-550b, nano-omni, muse-glimmer-30b, north-mini-code, laguna-xs-2.1) are all available: false and get redirected.
eco / Simple	moonshot/kimi-k2.7	google/gemini-2.5-flash-lite ($0.10/$0.40)	router-core ecoTiers.SIMPLE primary is nvidia/nemotron-3.5-lightning and its first fallback is nano-omni. Both are available: false now, so this takes the first live rung of that chain.
auto / Simple	moonshot/kimi-k2.7	google/gemini-2.5-flash ($0.30/$2.50)	Same as router-core tiers.SIMPLE primary.

All other routed ids are listed and available, so they did not change (deepseek-chat, gemini-2.5-pro, deepseek-reasoner, gemini-3.5-flash, gemini-3.1-pro, gpt-5.5, claude-opus-4.8, o3).

README examples, the free-model table and the routing table are updated to match. TestReadmeRoutingTableMatchesCode enforces the routing table. CHANGELOG has an Unreleased entry. VERSION is unchanged.

Guarding against regressions

• TestRoutingTableAvoidsRetiredIDs now covers every profile, not only free. Its denylist includes the ids above plus kimi-k2.5/k2.6/k2.7.
• New TestRoutingTableMatchesLiveCatalog is opt-in (BLOCKRUN_LIVE_CATALOG=1). It fetches /v1/models and fails if any routed id is missing or available: false, or if a free tier routes to a model that is not billed as free.

Checks

• go build ./...: ok
• go vet ./...: ok
• go test ./...: ok
• BLOCKRUN_LIVE_CATALOG=1 go test -run 'TestRoutingTable|TestReadme|TestFreeRouting' .: ok, against the live catalog
• integration_test.go (build tag integration, needs a wallet) was not run.

Note

The free tier now runs on a single model, an 11B vision model. Its quality on Complex and Reasoning prompts will be lower than nemotron's. That is the only free option the catalog currently marks available.

[VickyXAI]

Follow-up commit: the catalog's available flag is a live health signal. poolside/laguna-xs-2.1 flipped between true and false across five consecutive fetches. nvidia/llama-3.2-11b-vision was available on every fetch, which is why the free tier uses it. The router comment and the opt-in live test now say this, so an available:false failure gets refetched before anyone retargets.