Insights

Where Anomalies tell you something changed, Insights tell you something you could do: a concrete, usually dollar-quantified way to spend less. Each insight comes with an estimated monthly savings figure where one can be calculated, so you can prioritize the ones actually worth the time to act on.

These range across a few different angles: how your requests are shaped (response length, large or repeated inputs), whether a smaller, cheaper model could handle some of the work, and where your spend sits across platforms. A cheaper-model suggestion is a lead to test on a sample of real requests, not a guarantee the smaller model does the same job. When a vendor ships a newer model in the same family at the same or a lower list price, that shows up here too. The Insights tab shows exactly which ones apply to your account and why.

Some insights read details your providers report beyond the dollar total:

• Prompt caching. Models with large, mostly uncached input on a provider that bills cached input at a discount, with the saving worked out at that model's real cache prices. The estimate states its assumptions (how much of each prompt repeats), since that's the part only you know.
• Hidden reasoning. Keys where most of the output you pay for is a reasoning model's hidden thinking while the visible answers are short — usually a sign a lower effort setting would do.
• Longer responses. Keys where output is most of the cost and responses are getting longer.
• Long-context surcharges. Spend billed at a higher rate because prompts passed 200k tokens.

On Business and Scale, two billing checks are added:

• Billed vs list price. For each model, what the vendor charged against the same tokens at published prices, with the likely reason for any gap (priority processing, batch discounts, or a difference worth raising with the vendor).
• Why spend changed. Each week's change in spend split into how much you sent, which models it went to, and what the same models cost per token.
• One key carrying most of your spend. A leak or a stuck loop on that key would be most of your bill; splitting it by service or environment limits the damage.

For Anthropic and OpenAI, which report usage by the hour, TokenFront also looks for steady, around-the-clock traffic that never uses the Batch API. Traffic that runs evenly day and night is usually a pipeline or a scheduled job rather than people waiting on replies, and both providers bill batch requests at half price if the work can wait up to 24 hours.

On OpenRouter, where the provider that serves a model is a routing setting, TokenFront also checks whether another provider serves the same model for less. Only full-precision providers with at least 99% uptime over the last day are suggested — never a quantized build of the model.

On Business and Scale, the platform-concentration insight also says how many API incidents that platform's status page reported over the period.

Model-switch suggestions compare like with like: your exact tokens, cached input included, priced at both models' published prices and scaled by what you actually pay. A cheaper-model suggestion is never more than one size down from the same vendor. Each priced insight says which price data it used and when that data was last checked; newer models and long-context surcharges are tagged Market.

Insights are checked on a schedule rather than instantly, so a new one may take a little while to show up after a usage pattern starts. Think of it as a weekly-ish review rather than a live feed.

You can dismiss or snooze a single finding directly from the Insights tab if it doesn't need action right now, or turn off the underlying rule entirely from the Rules tab if it doesn't apply to how your team works.

Sign in to see your current insights: each one explains exactly what it found and what it would save.