Closed running code models, tools and agents
6 in the catalogue today, and the list grows as the market does. Every one with what it costs, who sells it and where it stands.
Somewhere for a model to run the code it just wrote, isolated from anything that matters. Billed by the second of compute and the memory held, not by tokens, so the cost follows how long the code runs. Look at cold-start time, what network access is allowed, and whether state survives between calls — an agent that must reinstall its dependencies every turn is an expensive agent.
Models whose weights are not published and which are bought through somebody's API. You get the maker's infrastructure, their scale and their uptime, and no way to run the thing yourself or to keep it if it is withdrawn. Most of the strongest models are here, so this is less a choice than a fact about the market; the choice is which seller you buy the same model from, and the catalogue holds their prices side by side.
- CompoundGroq1 selling
- Gemini 2.5 Flash-LiteGoogle3rd of 105· 3 boards$0.075→$0.3per Mtok in / out9 selling
- Grok 4.1 FastxAI21st of 34· 2 boards$5→$25per Mtok in / out2 selling
- Qwen3 Max ThinkingAlibaba$0.78→$3.9per Mtok in / out5 selling
- Sakana NamazuSakana AI$0.95→$4per Mtok in / out4 selling
- sandboxPerplexity$0.03per call1 selling