Enterprise AI spend control
Most companies send every AI request to the most expensive model available, because deciding case by case is hard. Krites makes that decision automatically — and keeps the expensive model for the work that actually needs it.
30-minute walkthrough on your own prompts. No data or API key needed to evaluate.
What it is
Every request passes through Krites on its way to a provider. It is read, checked against your policy, sent to a model, and recorded — so the same request is handled the same way every time, and you can see why.
A summary, a draft, a code review, a contract check — from the chat workspace or from your own application.
How difficult the task is, how much is riding on it, how sensitive the data is, and which models this team is permitted to use.
Routine work is handled cheaply. High-stakes work stays on the strongest model, determined by policy rather than by budget.
Each request is attributed to a team, a person and a use case as it passes through, so reporting requires no reconciliation later.
One real decision
“Summarize these 12 support tickets into three bullets for the standup.”
DeepSeek V4 Flash
Allowed by policy, capable of the task. Falls back to GLM 5.3 if it fails.
94× cheaper than premium
Legal, security, finance and executive work — anything scored high-risk — is never quietly moved to a cheaper model to save money. A team that is over budget gets slowed down or flagged, not given a worse answer.
For the engineers in the room
Open any of these for the mechanics.
A semantic classifier scores task type, complexity, risk and data sensitivity. It is negation-aware and clause-scoped, so “this is not a security issue” is not treated as security work. Metadata you send is trusted first; an optional LLM pass handles the ambiguous tail.
Org and team policy resolve into an allowed-model set the router cannot violate: whitelists, blocklists, sensitivity gating, premium access, premium limits and over-budget behaviour, applied in a fixed order. The classifier describes; the policy engine permits; the router optimises. No stage overrides the one before it.
Every decision carries a distinct fallback. Failed generations retry there; answers that fail quality checks escalate to a stronger tier automatically. Images force a vision-capable model, and PDFs, docx and spreadsheets are extracted to text and counted against the context budget before a tier is chosen.
Under the decision
Four more sit underneath it. Every request goes through all five, in this order, and they compound: a cheaper model thinking less about a shorter prompt whose repeated half was served from cache. Each layer is measured on its own line, so a saving that turns out not to be real cannot hide inside one that is.
The routing decision itself: the cheapest model on the ladder that the task's difficulty and risk allow, picked from the set your policy permits rather than the one the budget would prefer.
The same model can think hard or barely at all, and thinking is billed like any other output. The effort is proposed from the classifier signals routing already has, capped by your policy, then clamped to what that endpoint actually accepts — three layers, each of which can only lower it.
An agent resends its whole history every turn. A re-read of a file already in that history becomes a diff, an identical repeat becomes a one-line note, and old turns fold into a deterministic summary — without rewriting bytes the provider has already cached.
The stable part of a request — tool definitions, the system prompt, the conversation so far — is marked so the provider serves it from cache at a fraction of the input rate, with breakpoints placed to extend the cached prefix rather than rebuild it.
Failures retry on the decision's fallback and escalate a tier when an answer fails its quality check, all inside one usage record, so a retry can never bill twice. Cache savings, the premium paid to write those cache entries, and tokens compaction never sent are three different kinds of claim and are reported as three numbers.
Each layer has its own switch and its own measurement, so one can be turned off without disturbing the others — and none of them is allowed to report a saving the provider's own token counts do not support.
The bill
Here is the same request priced four ways — 2,000 tokens in, 600 out, at today's list prices.
A 60/30/10 mix across the three tiers — 83% below all-premium, with the hard work still on the best model.
Illustrative, at current registry list prices — not a customer result. Your real mix comes out of the eval harness on your own traffic before you commit to anything.
Attributed as it happens, so there is nothing to reconcile at month end.
Run rates projected to month end, broken into what is driving them such as more volume, longer answers, more premium use.
Premium models used on easy work, teams over budget, repeated prompts — each with a monthly cost and a recommendation.
Security & data
Routing and reporting run on details about each request — the kind of task, the model, the tokens, the cost. Storing the text itself is a separate setting, and it is off unless you turn it on.
Each prompt is stored as a one-way SHA-256 fingerprint. It identifies repeats and reuses cached answers, and it cannot be turned back into the original text.
Every request goes upstream with a no-training flag that cannot be switched off. You can additionally require zero-retention endpoints. Policy tightens, never loosens.
If no model meets your privacy rules, the request is refused with a clear error. It is never retried somewhere weaker to get an answer.
Who can see which workspace is decided on the server for every request, not in the browser — so access cannot be gained by tampering with the page.
Bring your own keys, add or remove models, reprice tiers in one file — routing, policy, cost maths and dashboards all read from it.
| Tier | Model | $ / 1M in | $ / 1M out | Best for |
|---|---|---|---|---|
| Simple | DeepSeek V4 Flash | $0.083 | $0.165 | Simple classification, short summarization, rewriting |
| Simple | GLM 5.3 Flash | $0.075 | $0.250 | Business drafting, multilingual work, standard tool use |
| Standard | GLM 5.3 | $1.40 | $4.40 | Coding, debugging, structured reasoning |
| Standard | GPT-5.6 Sol | $2.00 | $10.00 | Hard coding and debugging, long tool-call loops, analysis and planning |
| Premium | Claude Opus 5 | $5.00 | $25.00 | Complex reasoning, high-stakes analysis, legal/finance/security review |
A full chat workspace: conversations, attachments, web search, artifacts, citations, feedback. Routing happens underneath, so nobody has to think about which model they are on.
Users, teams, usage, use cases, router configuration, budgets and forecasts, audit log, and organisation settings.
Book a demo
Bring three or four prompts your team actually sends. We classify them live, show which model each one lands on, and put your current spend through the maths. The invite arrives immediately with a Meet link.
Calendar not loading? Open it in a new tab.