[VickyXAI]

What changes

This adds 17 ids to FREE_CHAT_MODELS in src/utils/constants.ts. Each is hidden from GET /v1/models, but the account price sheet (blockrun.ai/api/pricing) prices it at $0 / billingMode:"free". Until now an explicit call to any of them reserved the $5/$30 default, so an exhausted budget refused a free call. verify:prices had flagged each one.

The ids: nvidia/nemotron-3-super-120b-a12b-free, qwen3.5-397b-a17b, nemotron-ultra-253b, nemotron-super-49b, deepseek-v3.2, deepseek-v4-pro, deepseek-v4-flash, nemotron-3-nano-omni-nim, mistral-large-3-675b, qwen3-coder-480b, devstral-2-123b, glm-4.7, llama-4-maverick, qwen3-next-80b-a3b-instruct, qwen3.5-122b-a10b, seed-oss-36b, mistral-small-4-119b.

MODEL_TIERS.free (the routing order) is unchanged.

Bare-name safety

nvidia/deepseek-v4-pro has the same bare name as the paid deepseek/deepseek-v4-pro ($1.32/$3.96). Under the old map order the free id would have taken over bare deepseek-v4-pro, which means a $0 reserve for a paid call. To prevent that:

• BARE_TO_PREFIXED now adds free ids first, so a paid id with the same bare name always owns that spelling.
• New export FREE_SHADOWED_BY_PAID names the exception, which is currently only nvidia/deepseek-v4-pro.
• The uniqueness test allows only that kind of clash.
• A new test pins bare deepseek-v4-pro to the paid rate.

Sources

• blockrun.ai/api/pricing (2026-10-20): inputPrice/outputPrice 0, billingMode:"free".
• Gateway src/lib/models.ts: billingMode:"free", $0/$0 for 16 of the 17. deepseek-v4-flash is covered by the account sheet and by the MODEL_REDIRECTS entry in the chat route.
• Unpaid POST probes on blockrun.ai and sol.blockrun.ai (2026-10-20), using a ~3,000-char prompt with max_tokens 32. All 17 returned HTTP 200 with no payment header on both chains. The responding model was one of nemotron-3-super-120b, gpt-oss-20b, nano-omni or llama-3.2-11b-vision.

Checks

• npm run typecheck: clean
• npm test: 1423/1423 pass
• npm run build: ok
• npm run verify:prices: exit 0, 0 under-reserving. The catalogue sweep finds 0 gaps on Base, Solana and the account sheet. The 17 "billed $0 but FREE_CHAT_MODELS does not list it" notes are gone.

package.json is untouched.

Update: two DeepSeek mirrors dropped (2026-10-20)

nvidia/deepseek-v4-pro and nvidia/deepseek-v4-flash are removed from the free set. The gateway removed -flash in blockrun#848 (already live) and is removing -pro now. Once removed, both return an unknown-model 400, so the set now holds 15 account-sheet ids instead of 17.

The "paid wins bare spelling" rule stays: bare deepseek-v4-pro still reserves the paid deepseek/deepseek-v4-pro rate. FREE_SHADOWED_BY_PAID is now empty, and the tests pin that.

Expected verify:prices note: until the gateway change ships, the account sheet still lists nvidia/deepseek-v4-pro at $0. verify:prices therefore reports it as "billed $0 but FREE_CHAT_MODELS does not list it". That is an over-reserve note, not a failure. verify:prices still exits 0 with 0 under-reserving, and tests are 1423/1423.