Micro-model APIs
Lightning-fast
AI microservices
Too small for an LLM, too costly to host yourself: Chersus runs purpose-built micro-model APIs, sub-50ms, billed per input character, on sovereign EU infrastructure. Convenience without the bill, economics without the headcount.
Free up to 250K characters a month. No credit card.
curl -X POST https://api.chersus.com/v1/run \ -H "Authorization: Bearer chrs_live_YOUR_KEY" \ -H "Content-Type: application/json" \ -d '{ "input": { "type": "text", "value": "Contact Jane Doe at jane.doe@example.com about invoice #4821." }, "pipeline": [ { "id": "redact", "service": "governance.pii.redact" } ] }'
{
"trace_id": "chs_01J9X4Q7K2M8N3P5R6T7V8W9X0",
"status": "completed",
"output": { "type": "text", "value": "Contact [REDACTED] at [REDACTED] about invoice #4821." },
"results": {
"redact": { "kind": "transform", "status": "ok", "billed": "input", "changes": 2 }
},
"usage": {
"billed": [
{ "charge": "input", "unit": "characters", "quantity": 61 }
]
},
"audit": { "region": "eu-nl-1", "retention": "none", "total_ms": 26 }
}Compliance-clean licenses
Sovereign EU infrastructure
GDPR-compliant by design
< 50ms median latency
One service per job.
Governance
Strip or tokenize personal data before it reaches a log, a warehouse, or a downstream model. Detokenize on demand, from your own token map.
Security
Catch jailbreaks, prompt injection, and toxic content before they reach your model or your users. Each check returns a verdict and a score, ready to gate on.
Routing
Classify intent and detect language without hand-rolled keyword matching. Every message lands at the right step or queue.
Data
Turn free text into typed fields, parsed addresses, and ISO dates. Structured output your systems can insert without a parsing layer of your own.
Analytics
Score sentiment and traffic patterns across everything you route through Chersus. Structured scores come back, never prose.
Content
Tell AI-written text from human writing, watermark-aware where a provenance signal exists. A verdict and a score, ready to route into a review queue.
Command chaining
Clean and check in a single run.
One request · four services · four billed steps
Input
Contact Jane Doe at
jane.doe@example.com
40 chars in
governance.pii.redact
billed 40 chars
step 1
26 ms
parallel · 3 steps
security.toxicity.flag
billed 32 chars · 30 ms
security.jailbreak.shield
billed 32 chars · 18 ms
routing.language.detect
billed 32 chars · 9 ms
group 30 ms · max(30, 18, 9)
Result
Contact [REDACTED] at [REDACTED]
clean · 56 ms total
Billed 136 chars · €0.00544 past the free allowance · the same price as the four steps in sequence
Chain services in one request. Read-only steps can run at the same time on the same payload, so a fan-out of checks costs the latency of the slowest one, at the same price as running them in sequence. Every step bills the characters it processes, and a stop condition halts the run before the steps you do not need.
See chaining in the docs →Workflows: branch, loop, then route
Branch on a result and re-check with a cap: if a check flags, rewrite and try again, up to three times, then route. Same billing, same pipeline shape, planned on top of chaining.
Read the spec →Audit trace
Proof, never content.
{
"id": "redact",
"service": "governance.pii.redact",
"version": 2,
"model": "Roblox/roblox-pii-classifier-v2",
"input_sha256": "9f2c1e...a1",
"output_sha256": "6b7d44...0f",
"duration_ms": 26
}Every run returns an audit trace: SHA-256 hashes of the input and output, never the content itself. You can prove what ran, with which model and version, without exposing what was in it. You keep the receipt. We keep nothing.
Sovereign by construction.
Every request runs on sovereign EU infrastructure, across 26 cities, and never leaves European jurisdiction. Your input lives in memory for the duration of the run. We never persist, log, or train on it. When the request ends, the data is gone.
We act as processor under Art. 28 GDPR, with a simple DPA and a really short subprocessor list.
- Strictly European stack
- Hosted in the European Union
- No third-party models
- Zero retention policy
Billed per input character.
Start for free
€0
The first 250K characters each month. Every service, at full quality.
No credit card.
Request accessPay as you go, starting from
€0.04
per 1K characters, per step
One rate up to 5M characters a month. Above that it falls 21.5% for every doubling of volume, and it never drops below €0.002.
Example monthly rates
| Volume | Rate per 1K chars | Bill |
|---|---|---|
| 100K | €0.000 | €0 |
| 1M | €0.040 | €30 |
| 10M | €0.031 | €306 |
| 100M | €0.014 | €1,398 |
| 1B | €0.006 | €6,260 |
One step per request assumed. The formula bills; this table illustrates.
The sweet spot.
Hyperscaler LLMs buy convenience with money. Self-hosting buys money with headcount. Both are correct on one axis and wrong on the other. The sweet spot, low cost with nothing to own, is unoccupied. That is the position Chersus holds.
Start redacting in minutes.
Create an account, copy your key, make your first run before your coffee cools.