Anomalies

An anomaly is TokenFront noticing that something about your spending pattern doesn't match what's normal for your account, without you having to define "normal" yourself. It compares each day against your own recent history, not a generic benchmark, so what counts as unusual is specific to how your team actually uses AI. Yesterday's complete total is the main check. Today is checked as it goes too:

• Anthropic and OpenAI report usage by the hour, so TokenFront knows how much of a normal day's usage has usually happened by now. If today is running well ahead of that pace, you're told during the day: "on pace for about $X today", a projection that's re-checked every sync.
• Other providers report by the day, so today is flagged early only once its running total has already passed what a whole normal day costs, and the alert says "so far today".

Either way, a check can only see usage the provider has already reported, which can trail what actually happened by minutes or by many hours. See How current the data is.

You're alerted once per anomaly. If the same one keeps going, its numbers update in place rather than sending you a new alert each day. If it gets worse (a warning becomes critical), you're alerted once more.

Broadly, it's watching for things like a sudden spend spike or a key that's gone quiet when it's normally active, plus a handful of similar checks and any plain dollar budget caps you've set yourself. Sign in to see the full list of what's currently on for your account.

It also watches the market for the models you actually use:

• Price changes. When a model's list price changes, you see what it means at your own usage: the monthly dollar impact, worked out from the tokens you sent it over the last 30 days. A new price only counts once TokenFront's price sources agree on it.
• Retirements. You're told 90, 30 and 7 days before a model you spend on is retired, with the newest model in the same family and what switching would cost at your usage. Retirement alerts are included on every plan, Free too.

And for how your prompts are behaving:

• Prompt cache regressions. If the share of a key's input served from the prompt cache drops sharply while its volume stays steady, a prompt change has usually broken caching. You see the before and after, and what it's costing per day. Needs a platform that reports cached input (OpenAI, Anthropic, Vercel AI Gateway).

• Growing prompts. If a key's average input per request has grown every week for three weeks, the way agents and chat loops do when they resend their whole history, you see the growth and its monthly cost. Needs request counts, which Anthropic, Fireworks and xAI don't report.

• Retry loops. If a key's requests jump well above its usual day while the output per request collapses, a client is usually retrying failed calls or an agent is stuck in a loop. Needs request counts, which Anthropic, Fireworks and xAI don't report.

On Business and Scale, a "spend is trending up" alert also says what changed: more requests, larger requests, a shift toward pricier models, or a higher cost per token on the same models. Business and Scale also flag premium-tier creep: a rising share of a platform's spend on priority or fast processing, which is billed above standard rates. And when a spend spike, request surge, retry loop or quiet spell overlaps an incident on the provider's own status page, the alert says so and links to it. Status pages are checked for OpenAI, Anthropic, Fireworks AI and Vercel AI Gateway; OpenRouter, xAI and AWS Bedrock don't publish one TokenFront can read.

Price changes and retirements are tagged Market, and the Market filter shows just those. Each one names the price data it came from and when that data was last checked.

What you can do with one

Every anomaly can be acknowledged (seen, no action needed), snoozed (hide it for a while), or resolved (done, dismiss it). Anything left untouched clears itself automatically after 48 hours, so alerts don't pile up indefinitely.

How sensitive it is

You control how twitchy this is, anywhere from "only tell me if something's seriously wrong" to "flag any deviation at all," plus the specific dollar and day thresholds behind each one. All of that lives in the Rules tab, right next to the detector it affects, with the exact numbers and what they mean spelled out there.

Sign in to see this for your account: the full list of what's currently on, how sensitive each one is, and how to change it.