Token prices keep falling. AI bills keep growing. Here are five levers that cut agent costs without asking you to accept worse output, ranked by how fast they pay off.
The paradox eating your AI budget
In 1865, the English economist William Stanley Jevons noticed something strange about coal. Steam engines kept getting more efficient, needing less coal for the same work. Britain's coal consumption exploded anyway. Cheaper energy multiplied the number of things worth doing with coal, and demand swallowed the savings.
Tokens are living the same story. The price of a single token has dropped more than 90% since 2023, and yet spending on large language models has roughly doubled since late 2025.¹ Apollo's chief economist Torsten Slok calls it "Jevons paradox in action": as tokens get cheaper, companies respond by running more AI agents, automating more workflows, and generating more code, so total spend climbs even as each unit of intelligence gets cheaper.²
The market just shipped its own answer to this problem, and it named the product after the paradox. Jev, a "decision model" that launched in September, handles the small judgment calls an agent makes constantly: which tool to use, which model fits the task, whether a result is good enough to act on. Instead of generating text, it returns typed decisions with confidence scores, for a fraction of what a frontier model charges to think out loud.³ Its premise is the same conclusion the paradox points to: a huge share of agent spend goes to tiny decisions, and tiny decisions shouldn't pay frontier prices.
That's infrastructure-level change, and it will take time to reach most stacks. You don't have to wait for it. Cutting an AI agent bill comes down to a handful of levers you can pull this month, none of which needs new infrastructure, and none of which asks you to accept worse work. We run agents in production at Vybe, we've pulled all five on our own stack, and the rest of this piece is the playbook, ranked by how fast each lever pays off.
Lever 1: Match the model to the task
If every agent in your stack runs on the most expensive model, nobody made that decision. A default did. Fixing it is the highest-leverage hour you'll spend this month.
Go agent by agent and ask what each task requires. Extraction, classification, routing, formatting: these are mechanical jobs, and mechanical jobs don't need frontier intelligence. The price gap between tiers is enormous. On Anthropic's public API pricing, Claude Opus 5 costs $5 per million input tokens, Claude Sonnet 5 costs $2, and Claude Haiku 4.5 costs $1.⁴ Same tokens, same provider, a 5x difference between one of the top tiers and the bottom one...
The math gets concrete fast: say an agent summarizes 2,000 support tickets a day on a frontier model. Move that task to a small, fast model, and that line of the bill drops by roughly 80%. This is just an example, and your task mix will differ, but the ratio is the one we find on Anthropic's public pricing page.⁴
Just make sure you don't swap models blindly, and make it empirical instead: run both models on a sample of real tasks, compare the outputs, keep the cheap model only where the work holds up, and keep a revert path for everything. If quality drops on a task, move it back the same day. A downgrade you can undo in a minute is a downgrade worth trying.
(P.S.: you can't imagine the number of people who use Fable 5 for simple tasks...or lesser models for intelligence-intensive ones)
Lever 2: Stop paying for useless runs
Model choice was the first lever, but the second one is often overlooked: run frequency. Every scheduled run pays full input-token price for its prompt, even when there's nothing for it to do.
We audited our own scheduled agents recently and the findings were ugly. One hourly cron was on track to cost about $961 a month. A sync scheduled every 15 minutes ran 96 times a day, and almost every run had nothing new to process. A few of those across a fleet of agents...and you're funding a lot of empty polling.
Then there's the weekend paradox: our runs cost more on Saturday than on Wednesday. Token prices were the same as any other day, but the mix of agents running was different. The heavy scheduled agents were the ones still working while everyone was away, so the average cost per run climbed right when attention was lowest. A wasted run on Tuesday afternoon gets caught by a human, but a wasted run on a Sunday morning is easily missed.
What worked for us:
-
Audit at the agent level. Walk the whole roster of scheduled tasks and ask what each run is for. Cron-by-cron reviews can miss the pattern.
-
Match frequency to event rate. A sync that checks for new records every 15 minutes should run every 15 minutes only if records arrive that often.
-
Add a no-op exit. The cheapest run checks for work, finds none, and quits before touching a premium model.
-
Slow down runs whose output waits for a human anyway. If nobody reads the result until Monday, running it hourly through the weekend buys nothing.
Every scheduled run should earn its tokens.
Lever 3: Triggers over polling
A poller is an agent that pays to ask "anything new?" over and over. Most of the time the answer is no, and you paid for the question anyway. A trigger inverts the flow: the agent runs when the event happens and pays nothing while it waits.
On the agents we converted from polling to triggers, the trigger setup ran at about 5% of the polling cost. That's one stack's number rather than a benchmark, but the direction is hard to argue with: polling pays for silence.
Conversion has real limits, though. Keep polling when there's no event source to subscribe to, when you need a guaranteed check at a fixed interval, or when the poll is already cheap and rare. This lever pays most where frequency is high and the hit rate is low, which is exactly where the no-op cron burn from Lever 2 tends to live.
Lever 4: Give batch work a cheap base model
Some workloads are batch by nature, and they deserve their own economics. Tagging an entire database of contacts, cleaning a backlog of records, classifying a quarter of tickets: high volume, mechanical output, errors you can catch with a spot check.
One categorization job we ran, tagging a whole contacts database, cost about $2 for the entire job. Two things make math like that possible. Cheap models have crossed the quality bar for mechanical output. And batch jobs are cheap to verify: sample 50 results, check the tags, and you know whether the run is good. Errors in a batch are cheap to find and re-run. Errors in a live agent conversation are expensive to unwind. Looks like Jev, that I wrote about in the intro, can help much with this lever.
Pay for reasoning where reasoning pays. Batch work needs volume, so give it the model priced for volume (and add the necessary guardrails before scaling like mad, but that's common sense eh)
Lever 5: Cut context, carefully
The newest lever is the least obvious: most agent cost is input cost, and a lot of that input is context the task never needed.
The industry is converging on the fix from several directions. A Spotify engineer recently described routing bulk file reads to a cheap worker model, so the expensive model never sees the bulk I/O. Tested against a Java monorepo, mean savings on bulk reads ran around 90%.⁵
We ran our own version of the experiment: feeding webpages to a cheap extraction model before the main model saw them cut combined model cost by 36% on our evals, with answer checks passing.⁶
The caveat matters as much as the savings. Context reduction helps when it preserves enough context to finish the task. We also tried giving an agent broader file-read ranges instead of targeted ones, and it backfired: the agent burned extra reads rediscovering dependencies a full read would have shown it, and the change cost more than it saved. Trim what the task doesn't need. Keep the whole of what it does. We'll publish the full experiment results, including where the backfire showed up, in a follow-up piece.
The short version is the same lesson Not everything should cost a token makes about deterministic work: send the cheap model what it needs, and keep the expensive model for the thinking only it can do.
A sidebar: don't call a model when a tool will do
One pattern cuts across every lever above. Agent runs carry hidden meta-work that never reaches the user. Conversation compaction is the clearest case: when a thread grows long, its history gets compressed so it keeps fitting in context. The default is to ask the main model to summarize its own conversation, which means your premium model re-reads the entire thread at full input price to produce scaffolding. A specialized cheap tool, or plain deterministic trimming (keep the system prompt and recent turns, drop dead tool outputs), does the same job for a fraction of the cost.
Same family: structured extraction, classification, formatting, any step where the output is mechanical. Every one of those is a model call doing a tool's job.
Make it a ritual
These levers are habits. Prices will keep falling and usage will keep growing, that's what the paradox promises, so the bills that stay flat are the ones somebody watches.
Once a month, sit down with three numbers: cost per run, cost per agent, and both week over week. Look for which agents grew, which runs had nothing to do, which model tiers got expensive, and where retries stacked up. An hour a month, and the bill never gets to surprise you.
And the honest close: efficiency invites consumption. Every lever above makes agents cheaper to run, which makes it tempting to run more of them. That's Jevons again. The ritual is what keeps the paradox from eating the savings. A cut that survives a quality check is the only cut that counts.
If your AI vendor still bills per seat while your costs scale with usage, that mismatch compounds every month. We broke down why per-seat pricing is about to break if you want the full argument.
Try Vybe
Every lever above gets easier when your platform shows you the money: per-agent cost attribution, cost per run, and run histories a human can read. That's how we found the $961 cron, and it's how we keep our own bill honest. Try Vybe, or browse the gallery to see what agents are already built.
Sources
- Fortune, "Tokens are getting cheaper, but companies are spending even more on AI as a result, top economist warns" (June 17, 2026): the price of a single token has dropped more than 90% since 2023 while spending on large language models has roughly doubled since late 2025, per the Silicon Data Token Expenditure Index. https://fortune.com/2026/06/17/why-is-ai-spending-increasing-as-tokens-get-cheaper-jevons-paradox/
- Torsten Slok, Apollo, "Cheaper Tokens, Bigger Bills" (June 12, 2026): "This is Jevons paradox in action. As tokens get cheaper, companies don't spend less but instead run more AI agents, automate more workflows and generate more code, pushing aggregate expenditure higher even as the unit cost of intelligence collapses." https://www.apollo.com/wealth/insights-news/insights/daily-spark/cheaper-tokens-bigger-bills
- TypeSafe, "Introducing System One Models & Jev" (September 15, 2026): a decision model that returns typed decisions with calibrated confidence scores instead of generated text, priced at a fraction of frontier-model rates. https://typesafe.ai/blog/introducing-system-one-models-and-jev
- Anthropic API pricing, checked September 21, 2026: Claude Opus 5 at $5 per million input tokens, Claude Sonnet 5 at $2, Claude Haiku 4.5 at $1. https://docs.claude.com/en/docs/about-claude/pricing
- "Portal by Spotify cut my Claude Code token usage by 90%", Spotify Engineering (September 2026): "Tested against a Java monorepo across four scenarios... Mean bulk-read savings were around a whopping 90%." https://engineering.atspotify.com/2026/09/portal-by-spotify-cut-my-claude-code-token-usage-by-90
- Vybe internal experiment, September 2026: controlled runs with extraction-worker costs included, all answer checks passing. Not a production-wide savings estimate.
