Decide what you are actually buying
The word API covers several different products here, and the price lists only make sense once you know which one you need. LiteLLM is a gateway: it sits in front of models you already pay other providers for, and you bring your own key. ComputePrices.com and Salad sell compute, so you pick the hardware and run the model yourself. Azure AI Foundry is a model endpoint, where the provider hosts the model and bills you for tokens. A gateway, a rented GPU and a managed endpoint cannot be ranked against each other on a single figure, and the cheapest headline in each group hides a different cost.
Price per token, and what replaces it
Across the providers we track, input pricing spans €0 to €1.32 and output pricing spans €0 to €5.26, in euros, which shows how much room your usage pattern has. The three product types price differently. LiteLLM self-hosting is free, comes with a free tier under an open-source licence, and its Enterprise plan is sold as an annual licence priced to the gateway's annual request capacity rather than per token. Salad offers a flat monthly subscription or pay per token, with no free tier. Rented hardware has no token price at all, only a card rate such as $0.015 from ComputePrices.com or $0.015 from Salad, and a card you leave idle still bills.
The billing period decides whether the rate is real
Read the period before the rate. ComputePrices.com bills hourly, Salad bills per second, and UCloud bills hourly on GPU hosts with packages sold on monthly terms. A per-second charge suits a job that starts and stops, while an hourly charge suits a server you leave running between calls. Do the same arithmetic on the data side: a UCloud package carries 200 GB/month data package, 30 Mbps peak bandwidth, and a service that streams long answers back will consume that faster than a short classification request, so the egress terms can outrun the compute price on a chat-heavy product.
Limits, quotas and context window
Context window is the cleanest hard limit. The span across the data is 8,000 to 10,000,000, and Azure AI Foundry's published maximum is 1,050,000, so document-heavy work is one area where a managed endpoint can outrun a smaller model. Quotas are a separate story. ComputePrices.com grants 750 requests/day on a free tier model, enough to test a pipeline but not to carry production traffic, while Azure AI Foundry offers USD 200 in Azure credits to new customers. On a gateway, quota lives with each upstream provider, and LiteLLM can route across 140 of them, which is a resilience argument as much as a breadth argument.
Features and hardware shape the work you can do
Check interface compatibility first, because it decides how much code you rewrite. LiteLLM and Salad both offer an OpenAI-compatible API, so moving between them can mean changing a base URL instead of rebuilding an integration. Azure AI Foundry lists gpt 6 astra, gpt 6 luna and gpt 6 sol for sale directly, spanning text, embedding, image, audio and video, which lets a single provider cover a whole product surface. On rented hardware, compare the card lists directly: H100 SXM, A100 SXM and H200 against NVIDIA RTX 3060, NVIDIA RTX 3070 and NVIDIA RTX 3070 Ti. Then look at the interconnect, since Salad runs single-GPU nodes with no multi-GPU link, which puts tensor-parallel jobs out of reach there.
Support, contracts and compliance
Support is the checklist item that changes how an outage feels. LiteLLM and Salad both give you email, a help centre and community channels; ComputePrices.com gives email and a help centre; Azure AI Foundry gives a help centre and community. The named contracting entity is worth reading before procurement signs: LiteLLM is Berrie AI Incorporated, founded in 2023 and headquartered in the US; Salad is Salad Technologies, Inc., founded in 2018 and spread across from 190 locations; ComputePrices.com is Lansky Tech. If a client security team or a regulator sits in the deal, read the assurance lists: Azure AI Foundry publishes ISO 27001, ISO 27018 and SOC 1, covering FedRAMP, HITRUST and IRAP, while Salad holds SOC 2 Type I.
Run the comparison in the same order every time
Use a fixed sequence so the winner does not change with the question you ask last: product type, price and billing period, limits and quotas, features, then support and compliance. Score each provider in that order and the ranking does the rest. Start with the overall picture in the LLM APIs ranking, then read the individual verdicts for LiteLLM, ComputePrices.com, Salad, Azure AI Foundry and UCloud. If your workload sits near a borderline call, a Together AI or LLM Gateway review is the better second look.