uplinkswitch
Console
uplinkswitch

Features / Caching and context

Repeat answers at a tenth. Long chats that still fit.

Turn on the response cache with one header: an exact repeat is answered from cache at 10% of the original charge. Middle-out compression fits long chats.

Get a keyRequest exampleConsole docs

What it does

Opt-in response cache
One header, or a default on the key. Nothing is cached unless you ask.
A hit is a tenth
An exact repeat bills 10% of the original charge and shows HIT in x-uplink-cache.
Your organization only
The cache key includes your organization, so no other account ever reads your answers.
Middle-out compression
transforms: ["middle-out"], or context_overflow on the key, fits a long chat into the context window instead of failing it.
Shared prompt caching
On the Anthropic format a large system prompt that many accounts send can get a cache breakpoint; x-uplink-crowd-cache: off leaves a request out.
Readable headers
x-uplink-cache, -cache-ttl, -cache-age and -cache-source-id on cached answers; x-uplink-compressed on compressed ones.

Pricing · live catalog, 2026-11-04

RequestWhat happensCharge
First request (miss)6,000 in, 400 out on Claude Sonnet 5.5$0.0084
Exact repeat (hit)Answered from the cache, no model call$0.00084
Middle-out compressionDropped messages are never sent or billedNo fee

Request example

curl · cache one request
curl https://api.uplinkswitch.com/v1/chat/completions \
  -H "Authorization: Bearer $UPLINK_API_KEY" \
  -H "x-uplink-cache: on" \
  -H "x-uplink-cache-ttl: 3600" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-sonnet-5-5",
    "messages": [{"role": "user", "content": "What are your opening hours?"}]
  }'
Python · read the cache header
from openai import OpenAI

client = OpenAI(base_url="https://api.uplinkswitch.com/v1", api_key="sk-uplink-live-...")

raw = client.chat.completions.with_raw_response.create(
    model="claude-sonnet-5-5",
    messages=[{"role": "user", "content": "What are your opening hours?"}],
    extra_headers={"x-uplink-cache": "on"},
)
print(raw.headers.get("x-uplink-cache"))  # HIT or MISS
curl · middle-out
curl https://api.uplinkswitch.com/v1/chat/completions \
  -H "Authorization: Bearer $UPLINK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-sonnet-5-5",
    "transforms": ["middle-out"],
    "messages": [ ... a long chat ... ]
  }'
curl · defaults on a key
curl -X PATCH https://api.uplinkswitch.com/v1/keys/KEY_ID \
  -H "Authorization: Bearer $UPLINK_ADMIN_KEY" \
  -H "Content-Type: application/json" \
  -d '{"response_cache_ttl": 600, "context_overflow": "middle-out"}'
  • A hit answers with x-uplink-cache: HIT and the request id of the answer it replays in x-uplink-cache-source-id.
  • transforms is never sent on to the model; [] turns a key’s default off for one request.

Questions

The answer to a request from your organization, keyed by the model and the whole body (stream, user and metadata aside). An exact repeat within the time to live is answered from the cache. Chat completions, messages and embeddings, streamed or not; a streamed hit is replayed as the same events.

A tenth of the original charge. In the table, a Claude Sonnet 5.5 request that cost $0.0084 costs $0.00084 on a repeat, at today's catalog price.

Send x-uplink-cache: on, with x-uplink-cache-ttl from 60 to 86,400 seconds (300 unless set), or give the key a response_cache_ttl so every request on it is cached. x-uplink-cache: off skips a key’s default; x-uplink-cache-clear: true asks the model again and keeps the new answer. OpenRouter’s x-openrouter-cache headers work too.

When the prompt plus room for the answer is longer than the model's context window (1,000,000 tokens on Claude Sonnet 5.5), whole messages are dropped from the middle until it fits. System messages, the first user message and the latest turn are kept, and a tool call stays with its results. x-uplink-compressed says how many messages were dropped.

No. A request with a server tool, a redacted request and a fallback model’s answer are never kept, so a cached answer is always the model you named answering the body you sent.

More features

Every gateway feature · Console docs · Every model on one key · Hosted AI agents · Daily real-model check