Inference gets cheaper from both ends, Sol runs at $0.80
dax says a new DeepSeek rack beats every vendor on price/performance, while the Rails benchmark puts a full agent run at under a dollar.
dax posted that his team did their first test run on a new GPU rack with their new DeepSeek inference setup, and the price/performance is better than every vendor they have found. He added that there is still more optimization to do. That is the same rack we covered earlier, now with a first result attached.
Put it next to what the hosted models are charging and the shape of the year gets clear.
The $0.80 agent run
The Ruby on Rails account published Stage 3 of its Agents on Rails benchmark, run by evilmartians, covering seven new multi-step tasks spanning three to eight tickets each. These are not one-shot prompts. The tasks include building Rails apps from scratch and working through existing, sometimes messy, legacy codebases.
The best overall result went to OpenAI GPT-6.1 Sol, with 60.8% success at $0.80 and 23 minutes 35 seconds per run. Anthropic's Opus 5.5 at max effort was the most accurate at 65.3%, but it is also the most expensive and among the slowest. GPT-6 Astra at medium effort was fastest at 16 minutes 36 seconds and 54% success. The cheapest was GPT-6 Luna at $0.19 a run and 43.5% success.
DHH's read on the Sol numbers was blunt. Less than five percentage points off Opus, at $0.80 against $10.53.
Peter Steinberger, testing the same model, said Sol at 250 tokens per minute is a different game to Astra at roughly the same cost.
The business read
These are two different pressures on the same number. dax is pushing the floor down by owning the hardware and running open weights on it. OpenAI is pushing it down by shipping a model that does most of what the expensive one does for a fraction of the bill. If you priced an agent product against what a run cost you three months ago, that margin assumption is already stale, and nobody in these posts is claiming the fall is finished.
the price/perf we're getting is better than every single vendor that we've found. and there's still more optimization to do
Sol completes the tasks less than 5 percentage points off Opus, but at less than a tenth of the cost!!
Steal the numbers, not the vibes
One email a day. Under three minutes. Real MRR, real pricing, real launches from founders building with AI.