Main finding
Commercial AI APIs can power sophisticated systems, and in some workloads they are the better answer. The ATLAS claim is different: it is an integrated platform with financial and blockchain capability accumulated in advance, which must be demonstrated against those same APIs rather than assumed superior.
Benchmarks required before any claim
- Held-out directional accuracy on a fixed evaluation set, against a naive baseline and against the desk's current process.
- Execution and slippage error: modelled fill cost compared with realised fills on the same orders.
- Liquidity response under stress, compared with observed order-book behaviour in the recorded period.
- Cost per finished analysis, including integration, review time and human correction, not the API price alone.
- Time to add a new data source, measured in engineering days on both routes.
- Operating reliability: availability, reproducibility of the same question, and a stated failure rate.
- Security review against an agreed threat model, including key custody and data retention.
- Calibration of scenario ranges: does the stated interval contain the realised outcome at the rate claimed.
What supports it
- A general model plus retrieval reaches strong domain performance without retraining, which lowers the barrier to building specialized systems (Lewis et al. 2020).
- Tool use and multi-step acting extend what an API-based system can do operationally (Yao et al. 2022).
- Both routes carry integration, evaluation and operating cost. The comparison has to include all three, not the subscription price alone.
- Open-weight models are a third route, with hosting, tuning and governance cost of their own and no vendor support.
What this does not claim
- No head-to-head benchmark has been run on our tasks. The advantage is a hypothesis until it is measured.
- General models improve continuously and are available on commercial terms; a fixed licence can lag behind them.
Planning assumptions used
- Both alternatives are evaluated on the same tasks, data and acceptance criteria.
- Integration effort is costed for the licence route as well, not only for the API route.
Open before signature
- What is the evaluation set, and who signs off on it?
- What is the cost per finished task on each route?
- Which capabilities genuinely exist only inside the licensed platform?
Sources
Figures marked illustrative are planning assumptions. This note separates published evidence from project assumptions and does not present a supplier quotation, a valuation or a forecast of returns.