Model switch impact

What it does

Moving a workload to another model changes more than its price. Choose the model you use now and the one you are considering, and describe a month of the workload: what one request sends and returns, and how many requests there are. Or paste the usage object of one real response, and the tool reads its token counts for you.

It shows:

• The cost of the same load on each model, per request and per month, the difference and the difference over a year.
• Capabilities — the endpoints, streaming, image input, tool calling, JSON schema answers and reasoning — on each model, and which are lost or gained.
• Every request parameter the target rejects or ignores, as Salyro's catalogue records it, and which of them the current model accepts.
• The context window and the longest answer on each, and whether one request of your load fits the target.
• The tokenizers, and whether the same text is known to count the same on both.
• The lifecycle of each: active, deprecated with its retirement date, or not published by the provider.

Everything runs in your browser. A pasted usage object is read on the page; only its counts are kept, and nothing is sent anywhere.

How it computes

• One cost engine. The same month of requests is priced on each model by calculateCost, the function Salyro prices usage with, over Salyro's own price list, and the difference is the difference of the two answers. Price bands, cached input and cache writes are priced as each model's price list states them.
• Capabilities and parameters come from Salyro's catalogue — the one the gateway checks every request against. A parameter is rejected when the provider refuses a request that sets it, and ignored when the provider accepts it, answers at the model's defaults and says nothing. Through Salyro, a request that sets either kind is refused with a 400 naming the field, before it reaches the provider. Model parameter support lists every model.
• Context windows and tokenizers come from Reference Data, each with the page it was read from and the date.

What it does not know

• Whether the target counts your text the same. The load is counted in your current model's tokens. Where both tokenizers are recorded and the same (OpenAI's models on o200k_base, for example), the same text is the same count. Where either is not recorded — as for every Anthropic, Google and xAI model in the catalogue — it is unknown, and the target's own count, and so its cost, may differ.
• Answer quality and speed. The tool compares no benchmarks and no latency: no source it reads states them. Try the target on your own requests before you switch.
• A context window nobody recorded is unknown, not zero or unlimited.
• A model with no price for part of a request is not priced, never $0, and there is then no difference to state.
• Lifecycle a provider does not publish is shown as unknown, not as active.

FAQ

Do I need to change my code to switch?

Through Salyro, sending the same request to another provider is a change of the model id — see switching providers without changing your code. What this tool shows is what else to check first: a parameter the new model rejects or ignores is refused until you remove it, and a capability it lacks is refused too.

Why can the real cost differ from the figure here?

Because the target is priced at your current model's token counts, and a different tokenizer can count the same text differently. A longer or shorter answer changes the output tokens as well. Measure a sample of real requests on the target to confirm.

Why does Salyro refuse an ignored parameter instead of sending it?

Because an ignored parameter is a silent change: the request succeeds and the setting you asked for did not happen. A 400 naming the field tells you at once.

Sources

Prices come from each provider's pricing page, through Salyro's price list; the page names the revision and the date each was read. Capabilities, parameters and lifecycle come from Salyro's catalogue; context windows and tokenizers from the Reference Data listed under the page.