The same AI output, for about half the cost.
Most companies find they can get the same AI output for roughly half the cost. Not by using it less. By paying the price the work is actually worth.
How do you cut an AI bill in half?
Not by reducing usage. Camaze finds the places where the same output is being bought at the wrong price: frontier models doing work an open-weight or small model handles identically, prompts carrying context the task does not use, caching and batch pricing left switched off, agent loops retrying without a ceiling, workloads behind removed features, unused seats, and discount tiers you already qualify for.
Where the money actually goes.
Illustrative product view. Figures are examples.
Same shape every time.
What we found. What it costs. What to do. What you get back. No general advice, no maturity model, no scoring.
Wrong model for the job
**The largest finding, almost every time.** Systems get built against the strongest model and nothing ever moves them down.
- What we found
- A ticket triage workflow routes every message through a frontier model. 68% of those calls return a single label from a fixed set.
- What it costs you
- $19,400 a month, about 13% of total AI spend.
- What to do
- Route classification to a small model and keep the frontier model for the escalation path only. Both run against your existing eval set before the switch.
Context the task does not need
**Prompts accumulate and nobody trims them back.** Every extra token is paid for on every single call.
- What we found
- A document workflow attaches 14 retrieved chunks to every request. Measured against your own evals, answer quality is unchanged above 5.
- What it costs you
- $11,800 a month in input tokens that change no output.
- What to do
- Reduce retrieval depth to 5 and trim the system prompt of instructions the current model no longer requires. Configuration change, no rewrite.
Caching and batching left unused
**Both are configuration, not engineering.** Cached input and batch work are priced far below standard rates, and most eligible workloads use neither.
- What we found
- 78% of one workflow's input tokens are an unchanged system prompt sent on every call. Separately, a nightly enrichment job runs synchronously at full price despite having no latency requirement.
- What it costs you
- $13,800 a month across both.
- What to do
- Enable prompt caching on the static prefix. Move the enrichment job to the batch endpoint. Neither changes the output.
Agent loops and retries
**The fastest growing source of unexplained spend.** A retry that succeeds on the fourth attempt looks like a success everywhere except the bill.
- What we found
- A contract review workflow averages 11.8 model calls per run against a design target of 4. Retries account for 71% of its cost, and 3% of runs never terminate.
- What it costs you
- $16,200 a month, of which roughly $11,500 produces no result.
- What to do
- Set a per-run call ceiling and a cost ceiling. Add a terminal state for the failure path. Review the two tool calls responsible for most retries.
Workloads nobody has looked at since launch
**Something ships, works, and is never revisited.** A year later it serves a feature that was removed from the product.
- What we found
- Three workloads have had no code change in over nine months. One serves a feature retired in the last release. A dedicated self-hosted endpoint is running at 11% utilization.
- What it costs you
- $9,600 a month.
- What to do
- Retire the workload behind the removed feature. Consolidate the dedicated endpoint onto shared capacity. Re-benchmark the remaining two against current models.
Seats and duplicate tools
**Provisioning is easy. Reclaiming is nobody's job.** And two teams often pay for tools that do the same thing.
- What we found
- 84 assigned seats across two coding assistants have had no activity in 30 days. A third of the engineering organization has an active seat on both.
- What it costs you
- $6,100 a month.
- What to do
- Reclaim inactive seats and set an automatic reclaim rule at 45 days. Choose one assistant per team at the next renewal rather than running both.
Discounts you already qualify for
**None of them apply automatically.** Companies routinely pay list price for twelve months on volumes that qualified in month three.
- What we found
- Trailing twelve month volume on two providers supports a commitment at the next tier. Neither has been renegotiated since the original contract.
- What it costs you
- $14,200 a month against currently available terms.
- What to do
- Take the committed tier on the provider with stable volume. Open a renegotiation on the second using verified usage history as the basis.
Illustrative product views. Figures are examples.
Get your own list.
Connect read-only and the first pass runs against your last twelve months. Most companies have an itemized list with dollar figures inside a week.
The market moves faster than your budget cycle.
Optimizing once is a project. Staying ahead of it is the harder problem, because the ground moves constantly.
More detail
A cheaper model ships. A provider drops input pricing. Caching becomes available on a workflow that has been paying full rate for months. Your volume crosses into a committed tier. Every one of those is money, and none of it reaches finance in time to be planned.
New models, new prices and new discounts land every few weeks. You hear about them the week they land, priced against your own usage, in time to plan the change instead of reading about it afterward. That turns optimization from an annual cleanup into something finance can put on a cycle and push.
- New and updated models scored against your own workloads, volumes and evaluation results
- Price changes on models you already run, with the effect on your forecast applied automatically
- New capabilities such as prompt caching or batch pricing, matched to the workflows eligible for them
- Discount and commitment tiers as your volume reaches them, with a break-even calculation
- Deprecations and forced migrations flagged early, with the cost of each available destination
-
A cheaper model now covers your largest workload
Ticket triage runs 41M tokens a month on a frontier model. A newly released mid-tier model matches it on your own eval set.
$21,300a month -
Input pricing dropped on a model you already run
No change needed. The saving lands automatically. The forecast has been updated and plan variance is now favorable.
$4,900a month -
Prompt caching is now available on the document workflow
78% of that workflow's input tokens are an unchanged system prompt. Caching them is a configuration change, not a rewrite.
$8,700a month -
Volume now qualifies for committed-use pricing
Twelve months of usage supports a commitment at the next tier. Draft terms and a break-even are attached.
$14,200a month
Illustrative product view. Figures are examples.
Agents are where the new savings are.
An agent is a cost multiplier.
More detail
One user action becomes a chain of model calls, tool calls and retries, and the relationship between what the user asked for and what it cost is no longer visible anywhere.
The failure modes are specific and they are expensive: retries that eventually succeed and therefore never register as errors, loops with no terminal state, a frontier model used for every step in a chain where most steps are routing, and context accumulating across a long run until each call carries the entire history.
Camaze reports cost per run, calls per run against the design target, and the share of spend going to retries. That makes an agent a system with unit economics rather than a line item that grew.
- Cost per run, tracked against a target you set
- Retry share of run cost, alerted when it crosses a threshold
- Runs that never terminate, surfaced explicitly
- Step-level model selection, so routing steps are not billed at frontier rates
| Workflow | Runs | Calls per run | Cost per run | Spent on retries |
|---|---|---|---|---|
| Ticket resolution agent | 18,400 | 3.2 | $0.41 | 34% |
| Contract review workflow | 2,100 | 11.8 | $2.90 | 71% |
| Lead enrichment agent | 44,900 | 1.4 | $0.06 | 12% |
| Code review agent | 6,300 | 6.1 | $0.88 | 48% |
| Weekly report generator | 310 | 22.4 | $4.15 | 9% |
Illustrative product view. Figures are examples.
A ledger, not a report.
A list of recommendations that nobody works through is worth nothing.
More detail
Every finding has a dollar value, a workflow, an owner and a status, and it stays on the ledger until it is done or explicitly declined with a reason.
Finance sees the total identified, the total realized, and what is still outstanding. When a finding is implemented, the saving is verified against actual spend rather than assumed, so the realized number is defensible in a review.
- Optimizations ranked by annual savings, largest first
- Owner and status on each, with declines recorded and reasoned
- Realized savings verified against actual spend after implementation
- A running total of identified, in progress and realized
| Optimization | Workflow | Status | Savings |
|---|---|---|---|
| Open-weight model covers the classification step | Support triage | Done | $15,600 |
| Cache the system prompt, 78% of input tokens | Document workflow | In progress | $8,700 |
| Put a ceiling on the agent retry loop | Internal agents | In progress | $7,400 |
| Move batch-eligible work off synchronous | Data enrichment | Planned | $5,100 |
| Committed-use tier now earned | All providers | Planned | $14,200 |
| Right-size a GPU endpoint at 11% utilization | Self-hosted models | Under review | $6,900 |
| Reclaim unopened seats on two AI tools | Engineering | Under review | $3,900 |
| Total | $61,800 a month |
Illustrative product view. Figures are examples.
From connection to realized savings
Analyze
Read-only connection, twelve months of history backfilled, then a first pass across every workload. Most companies have an itemized list with dollar figures inside a week.
Work the ledger
Largest finding first. Configuration and commercial changes usually land the same week. Model and context changes go through your normal production review, validated against your evals.
Stay ahead
The radar keeps running. New models, price changes, new capabilities and discount tiers arrive as priced opportunities, so the ledger refills before the savings decay.
Cut the bill in half. Then spend it on more AI.
This is not an austerity exercise, and it should not be sold internally as one.
Spend less on the AI that is not working, so you can spend more on the AI that is.
Budgets get frozen when nobody can explain the number. Explain it, and the argument changes from defending the line to deciding where to reinvest.
Keep reading
Spend visibility
The collection layer underneath every finding: every provider, model and self-hosted workload on one cost basis.
Read moreAgent and workflow cost control
Retries, loops and runaway agents, and how to give an agentic system real unit economics.
Read moreAI savings calculator
Enter your monthly spend and provider count for an estimated savings range.
Try itQuestions people ask
How much can we actually save?
Most companies find they can get the same AI output for roughly half the cost. Where a company lands inside that depends on how long the systems have been running without review, how much of the workload sits on frontier models by default, and how much agentic work is in production.
You will see your own figure in the first week, itemized, with a dollar value and an owner against each finding. We would rather show you a specific number on your own spend than quote an average.
Does cutting cost mean using AI less?
No, and that is the point. Almost every finding on this page is about paying the correct price for work you are already doing: the right model for the task, context the task actually needs, caching that is available and switched off, retries that produce nothing, and discounts you have already earned.
Output volume is unchanged. The purpose is to free budget so you can run more AI, not less.
Do we have to change how anything works?
Mostly no. Caching, batching, seat reclamation, retiring dead workloads and taking a committed tier are configuration or commercial changes with no effect on behavior. Model routing and context reduction do change what runs, which is why Camaze reports them against your own evaluation set rather than asserting that quality will hold.
How does it know a smaller model would be adequate?
It looks at the shape of the work. A call that consistently returns one of a fixed set of short outputs is doing classification, whatever model is behind it. Camaze identifies candidates on that basis and quantifies the saving, then the change is validated against your evaluation set before anything moves. It proposes, it does not reroute traffic.
How do you keep up with new models and pricing?
Camaze tracks the model market continuously and prices every change against your actual workloads and volumes. When a cheaper model appears that covers something you run, when input pricing drops on a model you already use, when caching or batch pricing becomes available on an eligible workflow, or when your volume reaches a new discount tier, it appears as an opportunity with a dollar figure attached.
New models, new prices and new discounts land every few weeks. You hear about them the week they land, priced against your own usage, in time to plan the change instead of reading about it afterward.
Who does the work of implementing these?
Your teams. Camaze identifies the finding, quantifies it, names the likely owner and tracks it to completion. Configuration and commercial findings are usually same-day. Model and context changes go through whatever review process you already use for production changes.
Does this apply if we only use one provider?
Yes. Six of the seven finding types on this page are entirely internal to how your workloads are built and operated, and apply identically to a single provider. See single-provider visibility.