Track API-equivalent cost gaps by variation

[raychrisgdp]

Summary

This is the umbrella tracker for the live API-equivalent cost gaps in the dashboard. The current /api/models?days=30 snapshot still has 14 rows with positive token volume and missing or partial pricing, and the failures fall into four distinct buckets: subscription-aware pricing, alias mapping, source capture, and cost-component explanation.

Problem map

• Free-suffixed rows are being treated as zero or unknown when a real subscription/account price may exist — see Use API-key/subscription-aware pricing before treating free models as $0 #15.
• Alias resolution is missing for local/provider variants, so several high-volume rows cannot reach canonical pricing — see Add pricing alias registry for unmapped model IDs showing dash cost #16.
• GLM / Ollama rows need source tracing because cache exists in dashboard data, but the provenance of the pricing/capture path is still unclear — see Investigate missing prompt-cache accounting for GLM / Ollama subscription runs #31.
• The single cost number hides whether input, output, or cache is driving the estimate — see Show API-equivalent cost breakdown by token component #19.

Live evidence snapshot

Row	Tool / provider	Status	Why it matters	Child issue
nvidia/nemotron-3-ultra:free (nous)	Hermes / nous	partial	high volume, but still not fully priced	#15
gpt-5.3-codex-spark (openai)	OpenCode / openai	unpriced	pricing model ID exists, but estimate is still null	#16
mimo-v2.5-free	OpenCode / opencode	unpriced	free suffix plus no stable pricing path	#15 + #16
minimax-m3-free	OpenCode / opencode	unpriced	pricing model ID exists, but cost is still missing	#15 + #16
mimo-v2.5 (go)	OpenCode / opencode-go	unpriced	no canonical alias mapping yet	#16
nemotron-3-super-free	OpenCode / opencode	unpriced	meaningful token volume with no mapped price	#16
gpt-5.5-fast (openai)	OpenCode / openai	unpriced	pricing model ID exists, but estimate is still null	#16
gpt-5.3-codex-spark (openai-codex)	Hermes / openai-codex	unpriced	Codex alias needs canonical pricing coverage	#16
antigravity-gemini-3-flash (google)	OpenCode / google	unpriced	Google alias family needs canonical coverage	#16
glm-5.2 (go)	OpenCode / opencode-go	unpriced	cache is present, but capture/provenance still needs tracing	#31
nvidia/nemotron-3-ultra:free (openai-codex)	Hermes / openai-codex	unpriced	free suffix plus provider-specific routing	#15 + #16
antigravity-gemini-3.1-pro (google)	OpenCode / google	unpriced	another alias family member with meaningful usage	#16
nemotron-3-ultra-free	OpenCode / opencode	unpriced	free suffix with no stable canonical pricing path	#15 + #16
antigravity-gemini-3.1-pro-low (google)	OpenCode / google	unpriced	alias family member with meaningful usage	#16

Progress tracker

Child	Focus	Status
#15	API-key/subscription-aware pricing for free-suffixed models	open
#16	Pricing alias registry for unmapped model IDs	open
#19	Cost breakdown by token component	open
#31	GLM / Ollama prompt-cache and source-capture tracing	open

Current read

The work is not one bug. It is one pricing-resolution problem plus one provenance problem. Keep the umbrella issue as the index, and let the child issues own the concrete fixes.

Possible next split

If #16 keeps growing, the next clean split would be by alias family, not by individual row: OpenAI/Codex, Google/Antigravity, and opencode-go local variants. I do not think we need that split yet, but that is the first place I would cut if the registry starts becoming a grab bag.