Technical Context
I love such measurements more than any promotional benchmarks. Here, a person simply launched 13 parallel Sol Ultra sessions in fast mode and got a very down-to-earth result: minus 30% of the weekly limit in about half an hour, with 12 sessions managed to complete.
For me, this is not news about an "expensive model". It's news about AI automation in real research, where tokens become fuel, and poor AI architecture instantly turns work into a bonfire of budget.
According to available data, Sol Ultra has no separate surcharge specifically for fast mode. The main problem is different: the model likes to parallelize tasks via sub-agents, and the real consumption grows not by percentages, but by multiples. If your orchestrator is Sol Ultra, then every poorly configured fork easily inherits the same expensive model.
And this is where I braked. The discussion initially stated that parameters for sub-agents cannot be adjusted, but later clarified: in YAML, you can set role presets, model, and even separately specify a cheaper mode for some branches. So the issue is not just product limits, but how the orchestration itself is assembled.
If we take a rough estimate from open rates, Sol Ultra costs about $5 per million input tokens and $30 per million output tokens, and long context is even more expensive. With 13 parallel sessions with active branching, the bill can fly into hundreds of dollars per run. So the sentiment "we need a subscription for every 100 minutes of work" sounds hyperbolic, but essentially it hits the core nerve: an external researcher very quickly runs into economics.
Impact on Business and Automation
My first conclusion is simple: you cannot build a serious research stack on a single top model for all steps. I would keep the expensive reasoning model only for narrow decision points, while handing literature gathering, rough classification, fact extraction, and part of experiments to cheaper agents.
The second conclusion is less pleasant. Teams that already have near-unlimited access to tokens, data, and internal hypothesis validation pipelines win. Individual researchers and small teams lose if they go head-on without proper AI integration and sub-agent control.
And yes, I won't buy the thesis that "all science will collapse tomorrow into a couple of labs." In biomed, clinical, hardware, and everything that relies on real experiments, tokens are not the only bottleneck. But in fields with a huge search space, concentration will accelerate right now; it's visible to the naked eye.
We at Nahornyi AI Lab solve exactly such stories for clients: where to keep a strong model, where to cheapen orchestration, and where to completely replace an expensive pass with sensible automation with a human in the loop. If your research workflow already burns budget faster than it yields results, let's dissect the architecture and build AI solution development for your real load, not a fancy demo scenario.