Why AI costs spike, and how to stop the next one.

01Why AI costs keep rising when token prices fall.

The price of a token has fallen, while the number of tokens each task uses has risen faster. Agents make many calls per task, and each call re-reads the work so far, often after the model has reasoned at length before answering.

The FinOps Foundation reported in May 2026 that reasoning and agentic workloads consume five to thirty times more tokens per task than the equivalent chat interaction. Model changes can move the count too. Anthropic's pricing page says Claude 4.7 and later models use a newer tokenizer that produces about 30 percent more tokens for the same text, so an upgrade at the same list price can still raise the bill.

TakeawayA cheaper token does little for the bill when each task uses many more of them.

02What causes an unexpected AI cost increase.

Most spikes trace back to one of a short list of causes, and each leaves its own mark in the usage data. Knowing the mark tells you where to look first.

Two further causes make every one of these worse: spend that nobody attributes to a team, and alerts that arrive after the money is committed. The sections below take the causes in turn.

TakeawayEach cause has its own pattern in the usage data, so the pattern points to the fix.

03How agent loops and retries multiply AI spend.

An agent that retries a failing step pays for every attempt, and each attempt re-reads the whole task so far. A loop that runs overnight can spend more than the agent's normal month.

Loops usually start with something ordinary: a tool that returns an error the agent cannot parse, a check that never passes, or two agents handing work back and forth. Anthropic's cost guide for Claude Code lists scheduled tasks that fire while a session sits idle, sending the full context each time, and agent teams that use about seven times the tokens of a standard session in plan mode. OpenAI's Agents SDK takes a max_turns value and raises MaxTurnsExceeded when a run passes it, which is the guard many agents are deployed without.

Our guide to token budgets for AI agents covers turn limits, caps per run and what an agent should do when it reaches one.

TakeawayAn agent without a turn limit has no ceiling on what one bad run can cost.

04How reasoning tokens raise the bill.

Reasoning models think before they answer, and you pay for the thinking as output tokens even when you never see it. Output is the dearer kind of token, so a switch to a reasoning model can multiply cost per call while the visible answers look the same length.

• OpenAI's reasoning guide says reasoning tokens are not visible through the API, are billed as output tokens and occupy context. A model may generate from a few hundred to tens of thousands of them per request.
• On OpenAI, max_output_tokens counts reasoning, and a run can hit the cap before producing any visible answer, which pays for the reasoning and returns nothing usable.
• Anthropic's Claude Code cost guide says thinking tokens are billed as output and the default budget can be tens of thousands of tokens per request, depending on the model.
• Anthropic models numbered 4.6 and later keep earlier turns' thinking in context and bill it as input on later turns.

Anthropic reports how many billed output tokens were thinking in each response, so the share is measurable. A rise in output tokens with no change in request count usually points here.

TakeawayReasoning is billed as output, so set the effort level per task and watch output tokens per request.

05Why AI token costs climb as context grows.

Most chat and agent tools resend the whole conversation with each request, so the hundredth message pays for the ninety-nine before it. Context that is never cleared turns a cheap session into an expensive one.

Anthropic's Claude Code documentation says it sends the full conversation with every request, and that the first message after a break longer than the cache lifetime misses the cache and reprocesses the full context. Prompt caching softens the cost of a long prefix while it holds. When a prompt changes at the top, or a session sits idle past the cache lifetime, the same tokens are billed at the full input rate again. Price can also step up with length: on Anthropic's list, its smallest model charges a higher rate for prompts over 100,000 tokens.

TakeawayLong sessions pay for their whole history on every call, and a cache miss pays for it at full price.

06Why coding tools cause AI bill shock.

Coding assistants are sold per seat, then charge for usage above the seat's allowance. A few heavy users or one new agentic feature can push the total well past what the seat count suggested.

• GitHub says additional Copilot premium requests cost US$0.04 each, that each model carries a multiplier so some prompts deduct more than one request, and that counters reset on the first of each month at 00:00 UTC.
• Cursor's spend limits cap paid additional usage per period. When a limit is reached, additional usage stops for that user until the limit is raised or the next period starts.
• Anthropic puts average Claude Code spend at around US$13 per developer per active day across enterprise deployments, with 90 percent of users below US$30 per active day.

The averages hide the tail. A developer running long sessions on the largest model, with agent teams or scheduled loops, sits well above them, and on usage billing the overage arrives with the month's invoice.

TakeawaySeat prices set the floor of a coding tool's cost, and usage above the allowance sets the rest.

07Why alerts arrive after the money is spent.

Most spend alerts report spend that has already happened, and many controls only notify. Usage data and billing data also lag, so by the time an alert fires, the spike is hours or days old.

• OpenAI's spend alerts send a notification while API traffic continues. Hard limits return a 429 error, though OpenAI says enforcement is not instantaneous and recorded spend can slightly exceed the limit.
• Google Cloud says actual costs can take more than 24 hours to appear on billing reports, and that spend cap enforcement is not instant, with overages billed as normal.
• Anthropic's Usage and Cost API typically shows usage within five minutes of a request, though its cost report comes in daily buckets only. OpenAI's Costs API also returns daily buckets.

An alert also needs someone to send it to. When spend sits under one shared key or one organisation account, the alert reaches whoever owns the bill, who rarely knows which agent or team caused it. Our guide to tracking LLM costs per team covers how to attribute spend so each alert has an owner.

TakeawayAn alert helps only when it fires before the cap and reaches the person who owns the agent.

08How to diagnose an AI cost spike in a day.

Narrow the spike by time, then by model, then by key or project, then by token type. Each cut halves the search, and the token mix at the end names the cause.

1. Confirm the spike. Compare the provider's usage page with the invoice or cost report for the same days, and note which figures are estimates.
2. Find the start. Pull usage by day for the month, then by hour around the jump. Anthropic's usage endpoint takes one-minute, one-hour or one-day buckets.
3. Split by model. A jump on one model points to an agent or team that changed tier, or a new default.
4. Split by key, project or workspace. OpenAI's Costs API groups by project, API key and line item, and Anthropic's usage endpoint by API key, workspace and model. This names the agent or team.
5. Split by token type. More requests suggests a loop. More output per request suggests reasoning. More uncached input per request suggests context growth or cache misses.
6. Match it to a change. Check deployments, model or effort settings, new tools and new users on the day the spike began.
7. Check the coding tools separately. Their usage sits in each vendor's admin console, outside the API dashboards.
8. Contain it. Lower the limit on the key, project or workspace, add a turn limit, or pause the agent until the fix ships.
9. Write it down. Record the cause, the cost, the owner and the control that would have caught it sooner.

Date found:
Window (start and end):
Extra spend against the usual run rate:
Provider, model and key, project or workspace:
Agent or team, and its owner:
Token pattern (requests, output, uncached input):
Cause:
Containment and time applied:
Permanent fix and date:
Control that would have caught it sooner:

TakeawayCut by time, model, owner and token type, and the last cut usually names the cause.

09How to stop the next AI cost spike.

Give every agent and team a budget with a cap in the path its calls take, with an alert well before the cap and a named person who approves any increase. A spike then stops at a known amount instead of at the invoice.

1. Set a budget per agent and per team, in dollars per month, with a cap per run for agents.
2. Put the cap where calls pass, so it can stop a call before it is paid for. Provider limits make a good backstop under it.
3. Alert at about half and about 80 percent of each budget, sent to the agent's owner as well as finance.
4. Send requests for more budget to a named approver, with a reason and an expiry.
5. Treat a change of model, effort setting or tool list as a release, with a cost check before it ships.
6. Review cost per task for each agent every month, and retire the ones that cost more than they return.

Native limits differ by provider, and our comparison shows what each one lets you cap. To lower the run rate itself, our guide to reducing AI token costs covers model routing, caching and batch discounts.

Our AI spend management guide covers how to see the company-wide total these budgets add up to.

TakeawayA budget per agent with a named approver turns the next spike into a request.

10Common mistakes.

• Reading the invoice as the first signal. By then the month is spent.
• Running every agent under one shared key, so the spike has no owner.
• Switching to a reasoning model or a new model generation without re-measuring tokens per task.
• Setting spend alerts with no hard limit behind them, then assuming traffic stops at the alert.
• Raising a limit to clear a block without finding the cause, so the same loop runs again next month.
• Leaving coding tools out of the AI budget because they are bought as seats.
• Cutting usage across the board after a spike, when one agent caused it.