Up to your reserved tokens per minute, every request is already paid for. Past it, requests queue or bill at the Sync rate.

One flat monthly rate for the throughput you reserve.

RESERVEDALREADY PAID FORRESERVED TOKENS / MINONE FLAT MONTHLY RATEOVERFLOW · QUEUE OR SYNC RATE

Model your inference costs.

Compare published token rates against your workload and cache profile.

Rates & assumptions

Choose a model

Across all requests, not one response.

Eligible input only. Example, not a forecast.

50%
0%100%

Assumptions

Assumes 4:1 input/output · 24h/day · 30 days.

Optional service target

Fleet output

Across all requests, not one response.

Fresh inputCached inputOutput

Public-rate estimate

$16,070

per month · $192,845 annualized

50% cache scenario

$2,074

lower per month in this scenario

Reference · 0% input cache hits$18,144
Scenario · 50% input cache hits$16,070
Keep the model. Lower the cost.

Final pricing follows your workload benchmark.

DeepInfra via OpenRouter · 2026-09-06 · Rates & assumptions

USD per 1M tokens: $0.09 fresh input · $0.05 cached input · $0.34 output. Route: deepinfra/turbo, FP4. Model pricing · Provider rate source

Costs use aggregate output TPS × active hours × 3,600 × 30 days, plus input at your selected ratio. Annualized values use twelve 30-day months. Both scenarios use the same provider’s rates and output volume. Bar segments represent fresh input (gray), cached input (green), and output (black).

OpenRouter publishes a cache-read rate for this provider endpoint. Its endpoint metadata reported implicit caching as unavailable at retrieval. Do not assume a cache hit. Apply this rate only to input confirmed eligible by the provider's cache behavior and request configuration. Cache hits are an editable assumption, not a prediction. No separate cache-write rate is published. Additional cache-write, storage, network, platform fees, and taxes are excluded, not assumed free.

Market-reference provider route, not an OpenRelay availability or price quote. These token-charge scenarios are not your actual bill, measured savings, or an SLA. Final pricing and service targets require a workload benchmark. sales@openrelay.inc

Three ways to serve.

Pay for the response time you need.

REQUESTSSHARED POOLRELAY

Real time · per token

Sync

Responses as they run, billed per token at the catalog rate.

  • Chat
  • Copilots
  • Voice agents
View models
SYNCQUEUEYOUR ENDPOINT

Queued · early access20% off Sync

Async

Requests queue and results arrive when capacity frees up, at 20% less per token.

  • Agent workflows
  • Ingestion
  • Bursty traffic
Request access
JSONLAS CAPACITY FREESRESULTS

24h window · 50% off

Batch

Submit JSONL. It completes within 24 hours at half the Sync rate.

  • Evals
  • Synthetic data
  • Backfills
See batch API

Start with the SLA that matters.

Not sure which to pick? Start with tokens per second. Let us profile your workload and lower the bill.

TOKENS / SECONDTARGET

TPSRecommended

Tokens per second

Set a sustained output rate for the workload you actually run.

Best starting point
REQUESTTTFTITL

TTFT + ITL

Latency

Set response targets for chat, copilots, and user-facing agents.

User experience
TOKENS LEFTDEADLINE

Tokens / window

Batch deadline

Set a token volume and the time by which it must finish.

Completion time

Where we look for savings.

We profile your workload and benchmark each approach against your SLA.

SHARED PREFIXCACHE HIT

Cache-aware routing

Context reuse

Test cache-aware routing and shared-prefix reuse.

ONE BATCH

Kernels · batching

Kernel + batch tuning

Benchmark kernels, quantization, and batch shape.

PREFILLDECODE

Prefill · decode

Fit the hardware

Match prefill and decode phases to measured demand.

The selected approach depends on model and workload behavior.

Let's find the expensive tokens.

Send your model, traffic shape, and target SLA. We will define a representative benchmark.