Authors
Contents
Find out what your AI bill could drop to
The usual way to cut an AI bill is to cut usage: fewer seats, stricter rules, people quietly going back to doing things by hand. There is a better way. The same work can cost much less depending on how it is sent. Here are the levers that matter, split by what they cut, and the public prices behind them.
Start from the price list
AI pricing is per token, and the spread between models from the same provider is wide. At Anthropic's standard prices, Claude Opus 5.5 costs $4 per million input tokens and $20 per million output tokens. Claude Sonnet 5.5 costs exactly half that. Claude Haiku 5.5 costs $0.10 and $0.50, one fortieth of Opus (Anthropic pricing page, accessed 15 October 2026). Most teams never look at that list. They use the default model for everything, which is often the most expensive one.
The bill is growing too. CloudZero's State of AI Costs 2025 survey of 500 companies put the expected rise in average monthly AI spend at 36 percent, from $62,964 to $85,521 between 2024 and 2025. Only 51 percent of companies in the same survey said they could confidently measure the return. So the money is going up and most people cannot see where.
Once requests flow through one proxy, the levers below become available. They fall into four groups: send fewer tokens, pay less per token, never pay twice for the same context, and pay half price for work that can wait.
Fewer tokens: trim what each request carries
A long agent session resends its history on every turn. By the fortieth turn, most of what is being paid for is old conversation that the model no longer needs in full. Trimming that history and the context attached to each request, and setting sensible limits on answer length where a short answer does the job, cuts the number of tokens before any price is applied.
This is careful work. Trim too hard and answers get worse. That is why we check quality on your own work before and after every change.
Never pay twice: caching repeated context
Most requests from a coding tool or an agent carry the same large block of context every time: the system prompt, the project instructions, the file tree, the last few turns. Providers offer prompt caching so that a repeated prefix is billed at a reduced rate instead of full price on every call. On Claude Opus 5.5 and Sonnet 5.5 a cache hit costs 5% of the standard input price (Anthropic pricing page). On OpenAI models, cached input is billed at 10% of the input price (OpenAI pricing page).
Proxen makes full use of the providers' prompt caching, so the same repeated prefix is cached whether it comes from Claude Code, Cursor or a script. On top of that, our own cache answers requests that repeat exactly, so they are not paid for twice. The person sending the request does not have to do anything, and the saving shows up by person and by team in the spend view.
Lower price per token: routing simple work to cheaper models
Not every request needs the most capable model. Reformatting a file, writing a commit message, summarizing a ticket, classifying an email: these are jobs where a mid-tier or small model gives the same result at a fraction of the price.
We design the routing rules with you, per use case, and you approve every one. The proxy can pick a cheaper or a stronger model per task, even inside one agent session, and the log shows which model answered each request. Your team keeps asking for help the way they always did. The proxy sends simple work down the cheap path and keeps the strongest model for the requests that need it. Quality is checked on your own work before a rule goes live. Combined with a model list per team, routing is often the largest single lever.
Half price when it can wait: batch
Some work is not interactive. Nightly test generation, bulk document tagging, weekly report drafts, evaluation runs. Both OpenAI and Anthropic give a 50 percent discount on batch jobs that can wait up to 24 hours (OpenAI Batch API and Anthropic Message Batches documentation).
Batch suits jobs that can wait for their results. A batch job submits its requests once and collects the results later, so it has to be written for that; a script that waits on each answer cannot simply be switched over. We help you pick the jobs that fit and move them. Half price for the same tokens, in exchange for patience.
How the levers stack
The levers multiply rather than add, because each one works on what is left after the one before. On our illustrative example, built on public prices and a sample coding-agent session, the four together take a $4,000 monthly bill to about $1,500. That is an example, not a customer result, and the maths is shown on our AI cost page. Depending on the workload, the saving can reach up to 70-80%. Teams that already use cheaper models, or have caching set up well, save less.
Caps stop the surprises
The levers above cut the normal bill. Hard caps protect you from the abnormal one. In March 2026 The Register reported a single leaked Gemini API key that ran up $82,314 in 48 hours, against a normal bill of about $180 a month. A per-person key with a hard cap would have stopped at the cap. A raw provider key in a shared document has no cap at all.
Caps also make budgets real. A fixed monthly allowance per person, with an alert at 80 percent and a stop at 100, means the number finance signed off is the number that gets spent.
Measure before you cut
None of this works without a per-person, per-team view of spend. The order that works:
- Get every request through one address so it is counted.
- Look at a month of spend by person, team and model.
- Trim history and context, and fix caching. Both are cheap to try.
- Write routing rules for the top three request types by volume, and check quality on real work.
- Move anything non-interactive to batch.
- Set budgets from what you now know, not from a guess.
This is the work we do with you: we review the bill, sit with your engineers, build the proxy for your setup and keep tuning it as usage changes. Get a free AI bill review and we will show you where your money leaks and the saving to expect.
