Open model inference,
engineered for production
An OpenAI-compatible API you migrate to by changing three settings. Every request traceable, every cent explainable, upstream failures handled without you.
{
"model": "deepseek-ai/DeepSeek-V4.1-Flash",
"messages": [{ "role": "user",
"content": "Summarise this contract" }],
"stream": true
}- 1route-aTimed out at 800 ms · not billed
- 2route-bServed · switched before the first byte
Teams building on GridPort
- Northwind
- Lumora
- Parallax Labs
- Helioform
- Quanta Bio
- Orbitpay
- Velaworks
- Kitefin
- Monolith Data
- Brightwave
- Cinder & Co
- Tessellate
Model ecosystem
- DeepSeek
- Alibaba Qwen
- Z.ai
- Google Gemma
- Moonshot AI
- MiniMax
- Meta Muse
- Xiaomi MiMo
From 3 publishers, one catalog and one price source.
Design target for the time the gateway adds on top of the upstream first token. Measured latency per model is on each model page.
Monthly, for the serverless API, backed by service credits.
By email and in a response header, before any model retires.
Compatible is the floor, not the point
The real friction in migrating is parameter differences. The real friction in debugging is an upstream you cannot see. Both are built into the product.
Latency, tokens, cost and every routing attempt are recorded per request and kept for 90 days. Nothing about a bill is a mystery you have to reconstruct.
This was not a problem with your request.
Every request carries a Request ID. The failure reason, the suggested fix, the routing trail across attempts, tokens and cost sit in one drawer, and replay in the Playground.
Model allowlists, IP allowlists, key budgets, project budgets and three-stage balance alerts. Every block returns a specific error code; a non-streaming request already running completes, and a stream stops at its next funding segment.
- First token p50
- 186ms
- Throughput
- 212tok/s
- Success rate
- 99.8%
Sample chart: first-token latency stayed near 186 ms over the last hour; when an upstream slowed down, the route was switched before the first byte.
Model pages show first-token latency, throughput and success rate from the past hour. A failing route is switched before the first byte, and the switch is written to the headers.
Three lines change.
Your code keeps running.
The official OpenAI SDK works unpatched, and the migration assistant walks you through what to change. Parameters the platform does not know are ignored and named in the x-tf-ignored-params response header instead of failing the call. A parameter that needs a capability the model lacks, such as tools on a model without tool calling, is refused with 400 unsupported_parameter.
- 01/Point base_url at the platform
- 02/Swap in a platform API key
- 03/Change model to a platform model ID
import os
from openai import OpenAI
client = OpenAI(
base_url="https://gridport.ai/v1", api_key=os.environ["GRIDPORT_API_KEY"],)
stream = client.chat.completions.create(
model="deepseek-ai/DeepSeek-V4.1-Flash", messages=[{"role": "user", "content": "Hello"}],
stream=True,
)
for chunk in stream:
if chunk.choices and chunk.choices[0].delta.content:
print(chunk.choices[0].delta.content, end="", flush=True)
print()Featured models
Prices per million tokens. Pending models show official reference prices; connected models show GridPort billing prices.
Dedicated endpoints, custom models and compliance controls, enabled on request
Execution regions are written into the contract, and only routes with proof of location count. How it works →
- Dedicated endpoints
- Custom models
- SSO
- Audit logs
- Zero data retention
- Data residency
- Bank transfer & contract credit
Your first call in ten minutes
Sign up, create a key, copy the sample, get a response. There is no one you need to talk to.