Skip to content
prism

ASYNC LLM BATCH API. YOUR OUTPUT, YOUR PACE.

Send a batch.
Make room.

Summarize. Translate. Extract.
Text, Markdown, code or JSON.
A spending ceiling before you send.

Plan your pilot ↗
MINISTRAL 3 8B / BATCH API

$0.12 input · $0.12 output per 1M tokens

↓
01 / CLEAR THE WAY

Work that can wait.
Results you can track.

Send up to 100 text requests together.
Keep building while they run.
Collect each answer with its original ID.

A familiar request.
A durable place in the queue.
A cost you can see.

WHAT WILL YOU SEND?

Same batch. Different work.

Summarize & translate

Meeting notes → a Markdown recap.
Product copy → a translated description.

Free-text output · review tone and facts

Classify & extract

Support thread → a routing label.
Invoice text → fields in a JSON schema.

Optional structure · review extracted values

Draft & generate code

Return request → a draft reply.
A specification → a code snippet.

Text output · review before sending or running

JSONL carries the requests and responses. It does not require the model’s answer to be JSON. Text inputs only during the pilot.

02 / FOLLOW THE WHOLE EXCHANGE

One queue.
Many kinds of answers.

Upload your requests. Follow their progress.
Collect every result, with its original ID.

Open the workspace ↗
IN / YOUR APPLICATIONPython
import os
from openai import OpenAI

client = OpenAI(
    base_url=os.environ["PRISM_BASE_URL"],
    api_key=os.environ["PRISM_API_KEY"],
)

file = client.files.create(
    file=open("requests.jsonl", "rb"),
    purpose="batch",
)
batch = client.batches.create(
    input_file_id=file.id,
    endpoint="/v1/chat/completions",
    completion_window="24h",
    # Persist this key and reuse it for retries of this creation.
    extra_headers={"Idempotency-Key": "daily-extraction-2026-09-16"},
)

# Later, retrieve status and download results.
batch = client.batches.retrieve(batch.id)
THROUGH / PRISMAsynchronous
  1. 01

    Give it a place.

    Your requests are validated and saved before your job is accepted.

  2. 02

    Let it work.

    Follow your batch without keeping a connection open. Close the tab. Come back later.

  3. 03

    Make something of it.

    Download the responses and any per-request errors, ready for your next step.

Read the API guide ↗

24-hour processing window, best effort during the private pilot. No dedicated GPU reservation.

03 / SEE WHAT GOES INTO IT

Clarity,
down to the token.

01 / LEAN & MULTILINGUAL

Ministral 3 8B

Classification, everyday extraction and multilingual work.

Input / 1M
$0.12
Output / 1M
$0.12
Use this model ↗
02 / MORE ROOM TO THINK

Mistral Small 4

Structured extraction, summaries and more demanding language tasks.

Input / 1M
$0.12
Output / 1M
$0.42
Use this model ↗

USD pilot rates. Complete responses that pass the requested format checks are charged for actual tokens, including reasoning; failed requests are free. Factual errors remain billable. Unused reserved credit is released. Cached input uses the published input rate. Results are kept for 7 days. Live credit is operator-funded; checkout is test-only. Pilot limits ↗

Estimate token usage.Estimate your usage ↗
Estimated batch price$0.14

Based on your token volumes, before any applicable taxes.

Estimate your actual batch ↗

BATCH API QUESTIONS

Before your first request.

What is an LLM batch API?

It processes a group of language-model requests asynchronously. Submit your requests once, check their status later, and collect each response using its original ID. Prism has a 24-hour best-effort completion target during the pilot.

Does batch inference require JSON answers?

No. JSONL is the file format used to transport requests and responses. Each model answer can be free text, Markdown, code or JSON. Choose JSON Schema only when your workflow needs a checked structure. See request format examples ↗

Can I use the OpenAI Python SDK?

Prism supports a documented subset of the Files and Batches API. Set the Prism base URL and use your workspace API key. Streaming, tool calls and native media inputs are outside this pilot. Read the Python integration guide ↗

How much does a batch cost?

Prices depend on the model and actual token usage, including reasoning. Review a maximum charge before submitting. Failed requests are free; an accepted answer with factual errors remains billable. See token rates and pilot terms ↗.

START WITH ONE WORKFLOW

Plan your
first pilot.