We will look back at 2026 as the golden age of subsidized intelligence

For the same price as Netflix, you can get near unlimited access to frontier models
For $200/month you can employ a swarm of agents to build whatever you want
And enterprises are getting sweetheart deals to ensure provider lock-in
All subsidized by investors
But the tides are beginning to turn
Frontier labs are raising prices (modestly, for now) and beginning to move advanced capabilities behind usage based pricing models
The closer they get to post-IPO, the more they have to price for value, not adoption
What does this mean?
For companies
Smart companies will move more aggressively to post-train with open-weights models
The CFOs that allowed their orgs to tokenmax are seeing huge bills and will demand they get smarter with token spend
Modal just raised $355m and said this in their announcement:
"They're fine-tuning with their own data, running RL, and tuning inference for their own latency, throughput, and cost needs"
Why? Because open-weights models have reached quality levels that make this possible
They will still use frontier models a ton, but just not ALL workloads
It would be irresponsible for a company to not be exploring this approach
For consumers
We will enjoy benefits longer, but with increasing limitations
You'll still be able to get access to great models on ~$20/month plans, but:
> credits will start to get used up faster
> models will not perform as well as vs usage-based plans (i.e., you'll get served quantized versions and performance quality will always prioritize usage based requests)
> advanced capabilities will only be accessible via usage plans (i.e., things like Claude Design, etc.)
A small % of consumers who remain price insensitive will keep using the best of everything, creating an intelligence divide
For AI startups
If your business model involves inference packaged as a product, you need to plan for this now
Your pricing models will need to evolve
And you need to ensure your data capture can enable post-training on open-weights so you don't live or die by frontier lab APIs