⚡ China has a shadow market for AI APIs where people can get access to models like GPT-6 Astra or Fable 5.1 for as little as 10% of the official price.💰
We built Bestaiprice(dot)com - a catalog of ~200 AI API providers with some of the cheapest prices available.
We actually used many of these providers ourselves for our AI agents.
Then we decided to make the catalog public.
🕵️How do they get access to these cheap models?
A mix of shared software models and redistribution of unused tokens.
There are also gift and regional arbitrage schemes.
For example, through the Nigerian App Store, you can sometimes buy access to $200 worth of OpenAI credits for roughly half the price because of exchange-rate differences.
There are also promotional credits from Anthropic and OpenAI for companies - for example, $1,000 in Enterprise credits.
Microsoft and other OpenAI partners distribute access to OpenAI as well.🏷️
Then you have student quotas, multiple free-account registrations, promotional credits, and other unused allocations.
All of this creates a huge pool of cheap or free tokens that can be resold through API relays.
That’s how a model that officially costs 100% can sometimes end up costing ~10% on the shadow market.
We simply decided to map this market.
https://t.co/bjqXer2Yn1 ~200 providers for cheap AI APIs.
@SignalTech_ The real shift is in routing. If a cheaper model handles routine steps reliably, you can save the frontier budget for work that actually changes the result.
@theodorvaryag Visibility is half the battle. If you separate prompt, cached, and generated tokens per request, you can finally see where an agent is burning budget.
@damkina7 The big savings usually come from the architecture, not magic. Keep the strong model for hard steps, route routine work cheaper, then measure cost per successful task.
@calebfoundry The 65% token increase can erase a 20% price cut fast. For agents, I’d track cost per successful task and cache-hit rate, not just list price.
@0x_kaize The lower sticker price is only half the story. For agent workloads, cache behavior and success per task decide whether the model is actually cheaper.
@merccante That’s the part people miss. Most agent calls are decision overhead, so route the boring choices cheap and save the expensive model for work that changes the outcome.