SKTTECH skttech.io

SKTTECH · REGISTER 001 · SIGNAL OPERATIONS FOR AI TRAFFIC

Every model.
One endpoint.

ONE API · 400+ MODELS · 60+ PROVIDERS · ROUTED IN <3MS

++++
CH 01 · API CONSOLE · LIVE
claude 187ms deepseek 201ms qwen 164ms
REQ → TOK₁ 187 ms
SEC 02 · DEFINITIONS · MODULES 1–6 · EACH IN PRODUCTION

Six levers, all interlocked

MOD 02-ASmart Routing

route

price, latency, throughput — name the axis, we solve for it.

routing decision <3ms :nitro sorts by throughput p50 187ms :floor sorts by price $/M
"model": "qwen/qwen3-max:nitro"
"model": "deepseek/deepseek-v3.2:floor"

APPEND A SUFFIX — THE ROUTER HONORS IT ON EVERY CALL

MOD 02-BNo Lock-in

move

the exit is part of the architecture.

models 400+ providers 60+ any swap one string
"model": "openai/gpt-5.1"
"model": "deepseek/deepseek-v3.2"

LEAVING IS THE SAME ONE-LINE DIFF THAT BROUGHT YOU IN

MOD 02-COne Contract

agree

one schema, one invoice, four hundred vocabularies.

interface /chat/completions balance one, all vendors monthly fee none
base_url = "https://api.skttech.io/v1"
# the entire migration

YOUR SDK ALREADY SPEAKS THIS DIALECT — SO DO THE FRAMEWORKS

MOD 02-DObservability

observe

every request leaves an itemized trail.

per-request model · cost · latency rollups key · app · model export all of it — yours
"cost_usd": 0.000264,
"p50_ms": 187, "tokens": 36

THE DASHBOARD ANSWERS “WHAT DID THAT COST?” FIRST

MOD 02-EFallback

orchestrate

plan B is compiled in, not paged in.

declaration models: [primary, …] reroute on failure <50ms billed attempts successful only
"models": ["anthropic/claude-sonnet-5",
  "deepseek/deepseek-v3.2",
  "qwen/qwen3-max"]

AN OUTAGE BECOMES A LOG LINE, NOT AN INCIDENT CHANNEL

MOD 02-FZero Retention

forget

the most careful storage is none at all.

prompts & completions never stored ZDR routing verified providers training on your traffic never
NO STORED SECTION — DATA NOT RETAINED

ZERO RETENTION IS THE DEFAULT, NOT AN ENTERPRISE UPGRADE

also in the contract — SSE streaming · tool calling · structured outputs · multimodal · prompt caching · provider preferences · team quotas & audit logs

SEC 03 · SYNOPSIS · THE WHOLE MIGRATION

One line changes. Nothing else does.

from openai import OpenAI

client = OpenAI(
-   base_url="https://api.openai.com/v1",
+   base_url="https://api.skttech.io/v1",
    api_key=os.environ["SKTTECH_API_KEY"],
)

resp = client.chat.completions.create(
    model="anthropic/claude-sonnet-5",
    messages=[{"role": "user", "content": "Hello"}],
    extra_body={
        # automatic failover, tried in order
        "models": ["anthropic/claude-sonnet-5",
                   "openai/gpt-5.1",
                   "deepseek/deepseek-v3.2"],
        # cheapest provider for the model
        "provider": {"sort": "price"},
    },
)

Change one line, swap one key — nothing else in your codebase moves. LangChain, Vercel AI SDK, and LlamaIndex come along without being told.

00

Sign up

Takes an email, not a card. Your account arrives with $1 of routing credit already on it.

01

Top up

Credits, not subscriptions. Major international cards — wire transfer and invoicing for enterprises.

02

Get your key & go

Point base_url at api.skttech.io/v1, drop the key in, deploy. The diff above is the entire migration.

SEC 04 · EXAMPLES · INTERLOCKING BOARD · SAMPLE DATA

Watch a request make up its mind

++++
ROUTING POLICY

A model going down should be a fact about the weather, not a fact about your product. The router treats it that way: noticed in milliseconds, rerouted in under fifty of them, billed only where it succeeded.

0 T
Tokens routed / week
0 %
End-to-end availability
0 +
Models live
0 +
Upstream providers

sample data · live metrics wire in at launch

SEC 05 · RANKINGS · MEASURED, NOT MARKETED

The table your traffic writes

BERTH 01Anthropic

anthropic/claude-sonnet-5

7-day Δ +12.4% · coding

tokens / week 412.5B blended $/M $4.80 p50 latency 187ms
BERTH 02DeepSeek

deepseek/deepseek-v3.2

7-day Δ +9.1% · reasoning

tokens / week 388.1B blended $/M $0.31 p50 latency 201ms
BERTH 03OpenAI

openai/gpt-5.1

7-day Δ +5.2% · general

tokens / week 356.9B blended $/M $2.95 p50 latency 175ms
skt@ops:~$ rank | grep
#modelcategoryctx tokens/wkblended $/Mp507-day
The ranks change every week. The contract that reads them never does. the full board — all 400+ models → sample table · top 20 ≈ 86% of weekly traffic · blended $/M weights input and output at observed traffic ratios
SEC 06 · BILLING & PRIVACY · BOTH ITEMIZED

The ledger and the config

LEDGERCredits, itemized

Movement ledger · SKTTECH

top-up (card)$10.85
processing fee (5% of credited + $0.35)− $0.85
credited$10.00

usage — AI inference + platform services
anthropic/claude-sonnet-5
  in 12,406 tok$0.037218
  out 1,842 tok$0.027630
deepseek/deepseek-v3.2
  in 48,003 tok$0.013441

balance$9.921711
HIDDEN FEES = 0
no monthly fee · no minimum · purchased credits never expire
Transparent usage-based pricing. Your bill is calculated from the AI inference your requests use, together with the platform services around them — routing, high availability, monitoring, billing, and developer infrastructure. Pay only for success. A failed attempt that triggers failover costs $0. One request, one line item, however many providers it crossed. BYOK. Bring your own provider keys under your own agreements; the platform meters a usage-based service fee for routing, failover, and observability. Enterprise. 99.95% SLA, SSO, DPA, priority lanes, invoices, wire transfer — and a human who has already read your procurement checklist.

SEALEDRetention, set to zero

# privacy.toml — the default, not an enterprise upgrade
[privacy]
retention = "zero"    # prompts & completions, kept for exactly nothing
logging   = "metadata" # model, tokens, cost, latency — no payloads · full list: privacy policy
training  = false     # your traffic trains no one's model
 
[byok]
storage   = "aes-256" # your keys, encrypted; unbind anytime
scope     = "yours"   # your keys, your data, your bill
ZDR routing guarantee. Flag a request zdr: true, and it routes only to providers with verified zero-retention terms, or not at all. Enterprise DPA. Signed data-processing agreement, regional routing pins, and the 99.95% SLA in writing.
Does it work with the OpenAI SDK I already use?

Yes — full protocol compatibility, not “mostly.” Point base_url at https://api.skttech.io/v1, swap the key, and set model to a vendor/model id like anthropic/claude-sonnet-5. LangChain, Vercel AI SDK, and LlamaIndex run unchanged, because to them this is just an OpenAI-shaped server.

How does pricing work?

Transparent usage-based pricing. Your bill is calculated from the AI inference usage of your requests together with the platform services we provide — routing, high availability, monitoring, billing, and developer infrastructure. Every request is metered and itemized in the console, with no hidden fees. On BYOK traffic, where your vendor bills you directly, the platform charges a metered service fee.

How much latency does the gateway add?

Under 40ms end to end, and the routing decision itself takes under 3ms. For streaming, first-token latency matches going direct — we open the upstream stream and relay it; we do not buffer. Fleet-wide p50 is 187ms, and we publish it.

What are the rate limits?

There is no fixed RPM or TPM on your key. You draw on the platform's aggregated capacity across 60+ providers. When one provider throttles, the router reroutes automatically. Enterprise plans add priority lanes that hold capacity for your traffic.

If a request fails over to a second model, am I billed twice?

No. Only the attempt that returns your response is billed. A provider that errors or times out costs $0, however far into the stream it got. One request, one line item — that is the contract.

What payment methods and paperwork do you support?

Self-serve top-ups take major international cards, each with an e-receipt. Enterprises get wire transfer, proper invoices, and a signed DPA. Purchased credits never expire; there is no monthly minimum.

Who operates SKTTECH?

The SKTTECH platform (skttech.io) is operated by SKTTECH PTE. LTD., a company incorporated in Singapore. The platform is independent — not affiliated with or endorsed by any model provider; model names appear here only to identify routing targets. The paperwork lives at the usual addresses: Terms of Service and Privacy Policy.

Models are weather. The endpoint is climate — build on the part that holds still.

SKTTECH(1) · END OF ENTRY · SEE ALSO: ROUTE(2)