Support replies
OpenAI · gpt-4.1
Estimated
$8,200 baseline → $6,100 · gap $2,100
Product
One control plane between your app and every AI model.
Send parezen/auto. Parezen sends each request to the cheapest model that still clears your quality bar, checks it on live answers, and counts every saving against your usual model and your provider bill.
Free, read-only, no card. Sign up in a minute.
Requested model Cheaper model that passed your quality bar
Routing decisions
Sample data
- Support repliesgpt-4.1 gpt-4.1-mini26% lower
- Docs searchclaude-sonnet claude-haiku22% lower
- Nightly evalgemini-2.5-flash gemini-2.5-flash-lite42% lower
- Coding reviewgpt-4.1 gpt-4.1-mini6% lower
Illustration on sample data. Percentages come from the sample month.
Built for teams spending on OpenAI, Anthropic, Google and Mistral
OpenAI
Anthropic
Google
Mistral
Providers Parezen is built to work with, not customers.
Four steps. One ledger.
The same loop for every provider and every feature.
Connect
Upload a usage export or connect read-only billing. Nothing in your traffic changes to start.
See
Spend by provider, model, feature, team and agent, in one place, with budgets and hard stops.
Cut
A smaller model, a cache or a shorter prompt, switched on only after a shadow test passes your quality bar.
Prove
Every saving is a ledger line against your own baseline, then matched to the provider invoice within 2%.
One product. No subscription.
Every feature on every plan: 8% when you add credit, nothing on your own keys.
- Included
Auto routing
Send parezen/auto: each request goes to the cheapest model your team approved, with fallbacks and a receipt on every response.
See pricing - Included
Quality bar
Test questions and live checks against your usual model. A model that drops below your bar is removed at once.
See pricing - Included
Spend and savings
Spend by team, key, model and task, budgets and alerts, the savings ledger and the match against your provider bill.
See pricing - Contract
Enterprise controls
SSO and SCIM, model and provider allowlists, zero data retention, EU or US region, invoicing and an SLA.
See pricing
See where the money actually goes.
Switch the breakdown between provider, model, feature and team. The same sample month stays in view, and every total adds up to the same bill.
- Read-only. No change to your traffic.
- Your own baseline, not an industry average.
- The month shown here is a sample company.
SpendBaseline
Show the table
| provider | Spend | Baseline |
|---|---|---|
| OpenAI | $10,600 | $13,000 |
| Anthropic | $4,200 | $5,400 |
| Google Gemini | $3,800 | $5,800 |
| Mistral | $3,000 | $4,000 |
Show the table
| model | Spend | Baseline |
|---|---|---|
| gpt-4.1 | $10,600 | $13,000 |
| claude-sonnet | $4,200 | $5,400 |
| gemini-2.5-flash | $2,100 | $3,600 |
| mistral-medium | $2,100 | $2,600 |
| gemini-2.5-pro | $1,700 | $2,200 |
| mistral-small | $900 | $1,400 |
Show the table
| project | Spend | Baseline |
|---|---|---|
| Support | $6,100 | $8,200 |
| Eng | $4,500 | $4,800 |
| Docs | $4,200 | $5,400 |
| Eval | $2,100 | $3,600 |
| Ops | $2,100 | $2,600 |
| Finance | $1,700 | $2,200 |
| Catalog | $900 | $1,400 |
Show the table
| task | Spend | Baseline |
|---|---|---|
| Chat reply | $6,100 | $8,200 |
| Code review | $4,500 | $4,800 |
| Retrieval | $4,200 | $5,400 |
| Batch score | $2,100 | $3,600 |
| Classify | $2,100 | $2,600 |
| RAG | $1,700 | $2,200 |
| Batch draft | $900 | $1,400 |
Show the table
| agent | Spend | Baseline |
|---|---|---|
| None | $15,000 | $19,800 |
| review-bot | $4,500 | $4,800 |
| eval-bot | $2,100 | $3,600 |
Ways to spend less
Each lever has a stop. Anything that can change answer quality waits for a passed shadow test.
- Lever 1
Right-size the model
Send simple requests to a smaller model that still clears the quality bar.
How it works - Lever 2
Cascade
Try a cheap model first and escalate only when a check fails.
How it works - Lever 3
Provider prompt caching
Reorder prompts so stable parts qualify for cached-token discounts.
How it works - Lever 4
Response cache
Repeat requests are answered from the cache: exact matches on any route you turn on, near-duplicates on routes you list as safe.
How it works - Lever 5
Prompt and context trimming
Flag duplicate system prompts, oversized context and repeated instructions.
How it works - Lever 6
Output control
Set length limits, structured output and reasoning effort per route.
How it works - Lever 7
Batch routing
Move async work onto discounted batch APIs.
How it works - Lever 8
Waste removal
Catch retry storms, duplicate calls, agent loops and stale test keys.
How it works - Lever 9
Host arbitrage
For an open-weight model, pick the cheapest healthy host.
How it works - Lever 10
Shadow testing
Replay a sample of real traffic and score it before any switch.
How it works - Lever 11
Distil or fine-tune
Advise a tuned small model where one high-volume task would pay off.
How it works - Lever 12
Commit advisor
Advise on provider commitments and credits from actual usage.
How it works - Lever 13
Cross-provider switch
A cheaper model from another provider you connected, only after a replay of your own requests holds quality.
How it works
Every saving has a line. Every line has a baseline.
- The baseline is the model you asked for, with no levers, at the same price version.
- Estimated, then realised on real traffic, then verified against the provider bill.
- The total is always the sum of its lines.
Estimated $2,900
Realised $1,700
Verified $2,000
Docs search
Anthropic · claude-sonnet
Realised
$5,400 baseline → $4,200 · gap $1,200
Nightly eval
Google Gemini · gemini-2.5-flash
Verified
$3,600 baseline → $2,100 · gap $1,500
Coding review
OpenAI · gpt-4.1
Estimated
$4,800 baseline → $4,500 · gap $300
Invoice Q&A
Google Gemini · gemini-2.5-pro
Realised
$2,200 baseline → $1,700 · gap $500
Catalog copy
Mistral · mistral-small
Verified
$1,400 baseline → $900 · gap $500
Ticket classify
Mistral · mistral-medium
Estimated
$2,600 baseline → $2,100 · gap $500
| Feature | Provider | Status | Baseline | Spend | Gap |
|---|---|---|---|---|---|
| Support replies | OpenAI | Estimated | $8,200 | $6,100 | $2,100 |
| Docs search | Anthropic | Realised | $5,400 | $4,200 | $1,200 |
| Nightly eval | Google Gemini | Verified | $3,600 | $2,100 | $1,500 |
| Coding review | OpenAI | Estimated | $4,800 | $4,500 | $300 |
| Invoice Q&A | Google Gemini | Realised | $2,200 | $1,700 | $500 |
| Catalog copy | Mistral | Verified | $1,400 | $900 | $500 |
| Ticket classify | Mistral | Estimated | $2,600 | $2,100 | $500 |
| September total | $28,200 | $21,600 | $6,600 |
Baseline $28,200Spend $21,600Gap $6,600
Connect what you already use
Read-only access. Connectors never call a model and never change your traffic.
Usage export (Excel, CSV or JSONL)
Any provider. Upload last month's export; no keys needed.
OpenAI Costs API
Read-only admin key. Daily cost by model.
Anthropic Usage and Cost Admin API
Read-only admin key. Daily cost by model.
AWS Cost Explorer
Read-only IAM user. Amazon Bedrock cost by day.
Azure Cost Management
Read-only app registration. Azure OpenAI cost by day.
Google Cloud billing export
Read-only BigQuery access. Gemini and Vertex cost by day.
Your data, your call.
Parezen runs the gateway. Metadata only by default, and your app can always reach your provider directly.
Hosted gateway
Change one URL to api.parezen.com and bring your own provider keys, which Parezen stores encrypted. Parezen runs the gateway, so there is nothing to install.
Metadata by default
Tokens, model, cost and latency. Prompts and replies are kept only for workspaces that turn content logging on, for the retention you choose.
Always a way out
If Parezen is unreachable or slow to answer, your app can send the call straight to your provider. Budget hard stops are never bypassed.
Parezen is new and pre-revenue: there are no customers yet, so no figure on this site is a customer result. Pricing is published. The gateway runs inside the Parezen API at api.parezen.com, served only over TLS. It calls each provider with your own key and records tokens and cost from the provider's own usage fields. Provider keys are stored with envelope encryption, shown only by their last four characters and never logged, and every key and organisation is rate-limited. Parezen accepts a provider key only while all three protections are on.